{
 "defaults": {
  "boxes": {
   "decode": {
    "line": "136.274 tokens a second: Ollama 0.32.13, Q4_K_M, 350 W (the card's stock limit), a 4,096-token window, 256 tokens out (RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench). It is the card list's NVIDIA GeForce RTX 3090 (24 GB) · gemma4:26b: our most-measured card and its most-measured model, the lower-middle of 4 runs at the stock limit (135.2–146.3).",
    "option": "rtx-3090/gemma4-26b/ollama",
    "ref": null,
    "value": "136.3"
   },
   "prefill": {
    "line": "Our NVIDIA GeForce RTX 3090 (24 GB) reading a 98,000-token document. Its article doesn't name the model, so the card list can't offer this figure.",
    "option": null,
    "ref": "r69",
    "value": "2271"
   },
   "ttft": {
    "line": "510.56 ms: mistral-nemo:12b-instruct-2407, Q3_K_M with an 8-bit cache, Ollama 0.32.13, 300 W, a 4,096-token window, a 519-token question, one request (The quantization ladder). It is the card list's NVIDIA GeForce RTX 3090 (24 GB) · mistral-nemo:12b-instruct-2407, measured in the same run as that pair's 74.9; five waits were measured, 0.48–0.55 s.",
    "option": "rtx-3090/mistral-nemo-12b-instruct-2407/ollama",
    "ref": null,
    "value": "0.51"
   }
  },
  "settings": {
   "answer": {
    "n": 3,
    "unit": "paragraphs"
   },
   "answer_defaults": {
    "paragraphs": 3,
    "tokens": 256
   },
   "prompt_tokens": 519,
   "text": "prose",
   "wpm": 238
  },
  "source_line": "Starting numbers, all measured on our NVIDIA GeForce RTX 3090 (24 GB): writing speed with gemma4:26b, the wait with mistral-nemo:12b-instruct-2407, and reading speed with a model its article doesn't name. Details in “Where these numbers come from”, below the playback."
 },
 "exhibits": {
  "a-laptop-asked-the-desktops-questions": {
   "published": "2026-09-22",
   "sections": {
    "the-tables": "The tables"
   },
   "title": "A laptop, asked the desktop's questions",
   "url": "https://research.strata2signal.com/a-laptop-asked-the-desktops-questions/"
  },
  "a-short-history-of-mistral": {
   "published": "2026-09-16",
   "sections": {
    "what-a-power-cap-actually-costs": "What a power cap actually costs",
    "what-we-measured-on-two-cards-in-one-box": "What we measured on two cards in one box"
   },
   "title": "A Short History of Mistral",
   "url": "https://research.strata2signal.com/a-short-history-of-mistral/"
  },
  "flash-next-why-the-first-word-comes-sooner": {
   "published": "2026-10-12",
   "sections": {
    "the-24-gb-desktop-card-both-engines-in-one-window": "The 24 GB desktop card: both engines in one window",
    "the-desktop-card-one-engine-a-night": "The desktop card: one engine a night",
    "the-first-word-the-laptop-the-iq3-s-file": "The first word (the laptop, the IQ3_S file)",
    "the-q2-0-pair-the-laptop": "The Q2_0 pair (the laptop)",
    "writing-the-answer-the-laptop-the-iq3-s-file": "Writing the answer (the laptop, the IQ3_S file)"
   },
   "title": "Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench",
   "url": "https://research.strata2signal.com/flash-next-why-the-first-word-comes-sooner/"
  },
  "home-inference-and-diffusion-rig-cost": {
   "published": "2026-09-13",
   "sections": {
    "what-kind-of-performance-to-expect": "What kind of performance to expect"
   },
   "title": "The cost to purchase and run a very capable home AI rig",
   "url": "https://research.strata2signal.com/home-inference-and-diffusion-rig-cost/"
  },
  "ollama-and-vllm-on-one-rtx-3090": {
   "published": "2026-10-07",
   "sections": {
    "one-request-at-a-time": "One request at a time",
    "requests-at-once": "Requests at once"
   },
   "title": "Ollama and vLLM on one RTX 3090",
   "url": "https://research.strata2signal.com/ollama-and-vllm-on-one-rtx-3090/"
  },
  "one-3090-against-one-3090-ti": {
   "published": "2026-09-18",
   "sections": {
    "the-two-models-want-different-caps": "The two models want different caps",
    "three-more-arms": "Three more arms",
    "what-a-power-cap-takes-and-what-each-card-does-at-its-own": "What a power cap takes, and what each card does at its own limit",
    "what-fits-and-what-the-cache-costs": "What fits, and what the cache costs"
   },
   "title": "RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench",
   "url": "https://research.strata2signal.com/one-3090-against-one-3090-ti/"
  },
  "the-free-speed-wasnt-free": {
   "published": "2026-08-22",
   "sections": {
    "the-day-the-box-got-faster": "The day the box got faster"
   },
   "title": "The free speed wasn't free — the runtime dividend, traced to one commit",
   "url": "https://research.strata2signal.com/the-free-speed-wasnt-free/"
  },
  "the-quantization-ladder": {
   "published": "2026-10-19",
   "sections": {
    "speed-and-memory": "What each step down bought: speed and memory"
   },
   "title": "The quantization ladder",
   "url": "https://research.strata2signal.com/the-quantization-ladder/"
  },
  "the-typist-and-the-developer": {
   "published": "2026-08-27",
   "sections": {
    "one-card-both-jobs-at-once": "One card, both jobs, at once"
   },
   "title": "The Typist and the Developer",
   "url": "https://research.strata2signal.com/the-typist-and-the-developer/"
  },
  "two-hours-on-battery": {
   "published": "2026-08-31",
   "sections": {
    "the-one-table-to-remember": "The one table to remember",
    "the-wait-prefill-is-the-real-cpu-tax": "The wait: prefill is the real CPU tax"
   },
   "title": "Two Hours at 12 tok/s — On Battery",
   "url": "https://research.strata2signal.com/two-hours-on-battery/"
  },
  "two-used-3080s-priced": {
   "published": "2026-09-17",
   "sections": {
    "at-a-glance": "At a glance",
    "how-fast-does-each-one-write": "How fast does each one write?",
    "how-much-can-each-one-hold": "How much can each one hold?",
    "power-heat-and-the-fifty-watts": "Power, heat, and the fifty watts",
    "the-full-concurrency-ladder": "The full concurrency ladder",
    "what-about-a-long-document": "What about a long document?",
    "what-the-cache-type-costs": "What the cache type costs"
   },
   "title": "Two 3080s against one 3090, and the cap decides",
   "url": "https://research.strata2signal.com/two-used-3080s-priced/"
  }
 },
 "generated_from": {
  "roster": "https://research.strata2signal.com/hardware/rows.json",
  "rows": 1054,
  "rows_json_sha256": "128a3b2630f5f60c906a5730b6276392538359e923239fda4753b9a6d8639130"
 },
 "left_out": [
  {
   "exhibit": "two-hours-on-battery",
   "figure": "11.0 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "11.1 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "12.7 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "13.3 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "13.6 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "14.6 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "15.4 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "2.7 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "24.7 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "3.3 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "31.5 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "4.4 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "5.0 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "5.2 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "6.0 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "7.1 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "7.5 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "9.1 tok/s",
   "input": "decode",
   "label": "decode, median of five scored runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-one-table-to-remember"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "11.4 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "14.1 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "14.1 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "16.0 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "22.2 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "24.6 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "25.6 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "26.4 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "3.7 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "35.4 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "4.0 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "4.9 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "6.9 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "8.1 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "8.5 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "82.4 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "9.9 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "two-hours-on-battery",
   "figure": "97.7 s",
   "input": "ttft",
   "label": "time to first token, median of five cold-prefix runs",
   "reason": "a processor-only machine",
   "rule": "R2",
   "rule_words": "not a graphics card",
   "section": "the-wait-prefill-is-the-real-cpu-tax"
  },
  {
   "exhibit": "home-inference-and-diffusion-rig-cost",
   "figure": "137",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-kind-of-performance-to-expect"
  },
  {
   "exhibit": "home-inference-and-diffusion-rig-cost",
   "figure": "53",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-kind-of-performance-to-expect"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "73.327",
   "input": "decode",
   "label": "decode once the window is full, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "79.740",
   "input": "decode",
   "label": "decode once the window is full, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "80.056",
   "input": "decode",
   "label": "decode once the window is full, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "the-typist-and-the-developer",
   "figure": "207.7 tok/s",
   "input": "decode",
   "label": "writing speed",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "one-card-both-jobs-at-once"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "61.5 tok/s",
   "input": "decode",
   "label": "decode once the window is full, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-about-a-long-document"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "72.8",
   "input": "decode",
   "label": "decode once the window is full, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-about-a-long-document"
  },
  {
   "exhibit": "home-inference-and-diffusion-rig-cost",
   "figure": "0.08 s",
   "input": "ttft",
   "label": "time to the first token",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-kind-of-performance-to-expect"
  },
  {
   "exhibit": "home-inference-and-diffusion-rig-cost",
   "figure": "0.30 s",
   "input": "ttft",
   "label": "time to the first token",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-kind-of-performance-to-expect"
  },
  {
   "exhibit": "home-inference-and-diffusion-rig-cost",
   "figure": "1.01 s",
   "input": "ttft",
   "label": "time to the first token",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-kind-of-performance-to-expect"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "30.04 s",
   "input": "ttft",
   "label": "time to the first token, seconds",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-about-a-long-document"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "46.64 s",
   "input": "ttft",
   "label": "time to the first token, seconds",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "what-about-a-long-document"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "2,270.547",
   "input": "prefill",
   "label": "prompt reading, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "2,549.966 tokens a second",
   "input": "prefill",
   "label": "prompt reading, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "2,628.756",
   "input": "prefill",
   "label": "prompt reading, tokens a second",
   "reason": "its article names no model a reader can pick",
   "rule": "R3",
   "rule_words": "no model named",
   "section": "three-more-arms"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "525.93",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "528.06",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "694.10",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "702.42",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "the-tables"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "1,068.577",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "1,641.068",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "2,244.823",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "20,935.24",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "3,333.177",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "417.849",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "5,365.722",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "822.939",
   "input": "ttft",
   "label": "the wait for the first token, 95th percentile, ms",
   "reason": "more than one request at a time",
   "rule": "R4",
   "rule_words": "more than one request at a time",
   "section": "requests-at-once"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "56.174 (56.039–56.374)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "72.159 (72.024–72.171)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-tables"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "102.5",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "115.2",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "13.0",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "13.993",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "14.2",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "17.5",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "17.681",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "172.325",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "189.155",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "19.4",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "19.620",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "20.7",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "21.257",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "26.6",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "38.576",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "69.8",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "76.337",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "81.452",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "87.1",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "92.571",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "the-free-speed-wasnt-free",
   "figure": "124.975",
   "input": "decode",
   "label": "decode at a deeper context tier",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-day-the-box-got-faster"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "721.95",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "744.92",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-tables"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1.06",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1.9",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1.92",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "11.5",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "12.1",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "12.14",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "125.6",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "130.90",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "138.02",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "15.84",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "16.87",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "19.1",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2.52",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "20.91",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "204.4",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "26.63",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "321.8",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "45.2",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "47.6",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "50.0",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "51.05",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "528.72",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "545.1",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "566.57",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "6.37",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "7.6",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "73.3",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "9.5",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a wait measured at a 4,038-token prompt, over 600",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-q2-0-pair-the-laptop"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "94.22",
   "input": "ttft",
   "label": "first word, seconds",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1,362.8",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1,539.1",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,535.0",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,655.1",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,671.6",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,705.0",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "226.1",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "232.3",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "234.9",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "242.3",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "244.9",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "255.1",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "4,908.1",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "5,143.9",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-desktop-card-one-engine-a-night"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "626.7",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "640.9",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "a long prompt or a filled window",
   "rule": "R5",
   "rule_words": "a long prompt or a filled window",
   "section": "the-first-word-the-laptop-the-iq3-s-file"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "71.48",
   "input": "decode",
   "label": "decode, tokens a second, split, seats stopped",
   "reason": "by its label, one model split over two different cards",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "what-a-power-cap-actually-costs"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "38.22",
   "input": "decode",
   "label": "decode, tokens a second, split, seats up",
   "reason": "by its label, one model split over two different cards, with other work on them",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "what-a-power-cap-actually-costs"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "78.118",
   "input": "decode",
   "label": "decode, tokens a second (the server's own clock)",
   "reason": "by its label, a second clock on a run its (scored) row already counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "one-request-at-a-time"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "85.553",
   "input": "decode",
   "label": "decode, tokens a second (the server's own clock)",
   "reason": "by its label, a second clock on a run its (scored) row already counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "one-request-at-a-time"
  },
  {
   "exhibit": "the-free-speed-wasnt-free",
   "figure": "139.55 (n=6, range 137.75–150.49)",
   "input": "decode",
   "label": "sealed bench median decode",
   "reason": "by row, Ollama 0.32.9, before the upstream fix its own exhibit traces (the after-upgrade reading is 202-214 on 0.32.13)",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "the-day-the-box-got-faster"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "126.7",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "by row, a rounded reprint of 126.650 in “RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench”; the more precise copy counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "the-full-concurrency-ladder"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "134.8",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "by row, a rounded reprint of 134.754 in “RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench”; the more precise copy counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "the-full-concurrency-ladder"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "136.1",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "by row, a rounded reprint of 136.139 in “RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench”; the more precise copy counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "how-fast-does-each-one-write"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "33.3",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "by row, a rounded reprint of 33.265 in “RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench”; the more precise copy counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "how-fast-does-each-one-write"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "47.8",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "by row, a rounded reprint of 47.838 in “RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench”; the more precise copy counts",
   "rule": "R6",
   "rule_words": "a reviewed exclusion",
   "section": "how-fast-does-each-one-write"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "126.12 tokens a second",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "another version of the same engine on this card (not stated; 0.32.13 is shown: the version is stated; the other rows' exhibit does not state one)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "what-we-measured-on-two-cards-in-one-box"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "136.49",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "another version of the same engine on this card (not stated; 0.32.13 is shown: the version is stated; the other rows' exhibit does not state one)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "what-a-power-cap-actually-costs"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "34.13 tokens a second",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "another version of the same engine on this card (not stated; 0.32.13 is shown: the version is stated; the other rows' exhibit does not state one)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "what-we-measured-on-two-cards-in-one-box"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "48.18",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "another version of the same engine on this card (not stated; 0.32.13 is shown: the version is stated; the other rows' exhibit does not state one)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "what-a-power-cap-actually-costs"
  },
  {
   "exhibit": "a-short-history-of-mistral",
   "figure": "50.02",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "another version of the same engine on this card (not stated; 0.32.13 is shown: the version is stated; the other rows' exhibit does not state one)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "what-a-power-cap-actually-costs"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "22.490",
   "input": "decode",
   "label": "writing speed, tokens a second",
   "reason": "another version of the same engine on this card (3cf03257; b11514 is shown: the current build, 452 commits after 3cf03257, per its exhibit)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "78.126",
   "input": "decode",
   "label": "decode, tokens a second (scored)",
   "reason": "another version of the same engine on this card (0.27.1; 0.30.0 is shown: the newer of the two versions measured (0.27.1 on 2026-09-28, 0.30.0 on 2026-10-01))",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "one-request-at-a-time"
  },
  {
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "85.216",
   "input": "decode",
   "label": "decode, tokens a second (scored)",
   "reason": "another version of the same engine on this card (0.32.13; 0.34.4 is shown: the newer of the two versions measured (0.32.13 on 2026-09-28, 0.34.4 on 2026-10-01))",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "one-request-at-a-time"
  },
  {
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "239.7",
   "input": "prefill",
   "label": "prompt rate, tokens a second, the engine's clock",
   "reason": "another version of the same engine on this card (3cf03257; b11514 is shown: the current build, 452 commits after 3cf03257, per its exhibit)",
   "rule": "R7",
   "rule_words": "another version of the same engine",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "155.231",
   "input": "decode",
   "label": "summed tokens/s",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "155.251",
   "input": "decode",
   "label": "tokens/s",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "155.288",
   "input": "decode",
   "label": "tokens/s",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "44.702",
   "input": "decode",
   "label": "tokens/s",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "44.897",
   "input": "decode",
   "label": "tokens/s",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "121.733 (121.722–121.762)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "155.712 (155.641–155.908)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "165.656 (165.545–166.175)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "44.946 (44.917–44.963)",
   "input": "decode",
   "label": "tokens/s (spread of three)",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "146.780, 19,428 MiB",
   "input": "decode",
   "label": "decode at the full window, and the memory it used",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-fits-and-what-the-cache-costs"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "159.410 tok/s, 21,008 MiB",
   "input": "decode",
   "label": "decode at the full window, and the memory it used",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-fits-and-what-the-cache-costs"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "126.650",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "134.754",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "145.817",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "146.996",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "147.201",
   "input": "decode",
   "label": "decode summed across conversations, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "three-more-arms"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "136.139",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "145.908",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "147.599",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-two-models-want-different-caps"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "147.823",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "33.265",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "42.969",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "47.838",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "53.354",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "56.936",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "450 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "124.7 tok/s, 19,415 MiB",
   "input": "decode",
   "label": "decode at the full window, and the memory it used",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-the-cache-type-costs"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "133.6 tok/s, 20,995 MiB",
   "input": "decode",
   "label": "decode at the full window, and the memory it used",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "what-the-cache-type-costs"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "125.6",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "how-fast-does-each-one-write"
  },
  {
   "exhibit": "two-used-3080s-priced",
   "figure": "33.1",
   "input": "decode",
   "label": "decode, tokens a second",
   "reason": "350 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "how-fast-does-each-one-write"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "132.94",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "191.65",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "203.72",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "208.97",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "336.23",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "371.01",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "372.92",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "401.70",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  },
  {
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "439.78",
   "input": "ttft",
   "label": "first token, ms",
   "reason": "95 W, the stock limit, was measured too",
   "rule": "R8",
   "rule_words": "not at the card's stock power limit",
   "section": "the-tables"
  }
 ],
 "limits": {
  "boxes": {
   "decode": {
    "hi": 2000,
    "lo": 0.5
   },
   "prefill": {
    "hi": 100000,
    "lo": 1
   },
   "ttft": {
    "hi": 120,
    "lo": 0.01
   }
  },
  "max_sets": 2,
  "settings": {
   "answer_paragraphs": {
    "hi": 5,
    "lo": 1
   },
   "answer_tokens": {
    "hi": 4096,
    "lo": 10
   },
   "prompt": {
    "hi": 131072,
    "lo": 10
   },
   "wpm": {
    "hi": 600,
    "lo": 100
   }
  },
  "texts": [
   "prose",
   "code"
  ],
  "units": [
   "paragraphs",
   "tokens"
  ]
 },
 "picker": {
  "cards": [
   {
    "name": "NVIDIA GeForce RTX 3080 (10 GB)",
    "order": 1,
    "slug": "rtx-3080",
    "stock_cap": null
   },
   {
    "name": "Two NVIDIA GeForce RTX 3080 (10 GB each)",
    "order": 2,
    "slug": "two-rtx-3080",
    "stock_cap": null
   },
   {
    "name": "NVIDIA GeForce RTX 3090 (24 GB)",
    "order": 3,
    "slug": "rtx-3090",
    "stock_cap": "350 W"
   },
   {
    "name": "NVIDIA GeForce RTX 3090 Ti (24 GB)",
    "order": 4,
    "slug": "rtx-3090-ti",
    "stock_cap": "450 W"
   },
   {
    "name": "NVIDIA GeForce RTX 5090 Laptop GPU (24 GB)",
    "order": 5,
    "slug": "rtx-5090-laptop-gpu",
    "stock_cap": "95 W"
   },
   {
    "name": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)",
    "order": 6,
    "slug": "rtx-pro-6000-blackwell-workstation-edition",
    "stock_cap": null
   }
  ],
  "options": [
   {
    "caps": [
     "not stated"
    ],
    "card": "rtx-3080",
    "engine": "not stated",
    "engine_version": "not stated",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "not stated"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "34.4",
      "max_shown": "34.4",
      "min": "34.4",
      "min_shown": "34.4",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r1",
      "refs": [
       "r1"
      ],
      "shown": "34.4"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": null,
    "notes": [
     "The model did not fit in the card's 10 GB and ran partly from system memory (the exhibit labels these rows “split with system memory”).",
     "One card carries figures whose record is not on this shelf. The RTX 3080’s readings are published — on a released page, in a released table, provable word for word against that page’s own row — but every one of them cites a bench file this hub does not publish, and the page itself says so. Its card prints the figures and says the same thing rather than implying a receipt that does not exist."
    ],
    "slug": "rtx-3080/gemma4-26b/not-stated",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "not stated"
    ],
    "card": "rtx-3080",
    "engine": "not stated",
    "engine_version": "not stated",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "not stated"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "5.13",
      "max_shown": "5.1",
      "min": "5.13",
      "min_shown": "5.1",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r2",
      "refs": [
       "r2"
      ],
      "shown": "5.1"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": null,
    "notes": [
     "The model did not fit in the card's 10 GB and ran partly from system memory (the exhibit labels these rows “split with system memory”).",
     "One card carries figures whose record is not on this shelf. The RTX 3080’s readings are published — on a released page, in a released table, provable word for word against that page’s own row — but every one of them cites a bench file this hub does not publish, and the page itself says so. Its card prints the figures and says the same thing rather than implying a receipt that does not exist."
    ],
    "slug": "rtx-3080/mistral-small3.2-24b/not-stated",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "250 W each, 500 W for both cards",
     "300 W each, 600 W for both cards",
     "not stated"
    ],
    "card": "two-rtx-3080",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "250 W each, 500 W for both cards",
       "300 W each, 600 W for both cards",
       "not stated"
      ],
      "complete": 6,
      "how": "anchor",
      "max": "124.7",
      "max_shown": "124.7",
      "min": "113.8",
      "min_shown": "113.8",
      "n": 6,
      "r8_left_out": 0,
      "ref": "r3",
      "refs": [
       "r4",
       "r5",
       "r3",
       "r6",
       "r7",
       "r8"
      ],
      "shown": "115.4"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": null,
    "notes": [],
    "slug": "two-rtx-3080/gemma4-26b/ollama",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "250 W each, 500 W for both cards",
     "300 W each, 600 W for both cards"
    ],
    "card": "two-rtx-3080",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "250 W each, 500 W for both cards",
       "300 W each, 600 W for both cards"
      ],
      "complete": 2,
      "how": "anchor",
      "max": "44.7",
      "max_shown": "44.7",
      "min": "44.6",
      "min_shown": "44.6",
      "n": 2,
      "r8_left_out": 0,
      "ref": "r9",
      "refs": [
       "r9",
       "r10"
      ],
      "shown": "44.6"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": null,
    "notes": [],
    "slug": "two-rtx-3080/mistral-small3.2-24b/ollama",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-3090",
    "engine": "Ollama",
    "engine_version": "0.34.4",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "85.007",
      "max_shown": "85.0",
      "min": "85.007",
      "min_shown": "85.0",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r11",
      "refs": [
       "r11"
      ],
      "shown": "85.0"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.441866",
      "max_shown": "0.44",
      "min": "0.441866",
      "min_shown": "0.44",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r12",
      "refs": [
       "r12"
      ],
      "shown": "0.44"
     }
    },
    "label": "Gemma 4 12B · Ollama",
    "model": "Gemma 4 12B",
    "narrowed_to": null,
    "notes": [
     "The wait is the slow end: 95 in 100 first words came sooner."
    ],
    "slug": "rtx-3090/gemma-4-12b/ollama",
    "stock_cap": "350 W",
    "version_why": "the newer of the two versions measured (0.32.13 on 2026-09-28, 0.34.4 on 2026-10-01)"
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-3090",
    "engine": "vLLM",
    "engine_version": "0.30.0",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "82.693",
      "max_shown": "82.7",
      "min": "82.693",
      "min_shown": "82.7",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r13",
      "refs": [
       "r13"
      ],
      "shown": "82.7"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.212999",
      "max_shown": "0.21",
      "min": "0.212999",
      "min_shown": "0.21",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r14",
      "refs": [
       "r14"
      ],
      "shown": "0.21"
     }
    },
    "label": "Gemma 4 12B · vLLM",
    "model": "Gemma 4 12B",
    "narrowed_to": null,
    "notes": [
     "The wait is the slow end: 95 in 100 first words came sooner."
    ],
    "slug": "rtx-3090/gemma-4-12b/vllm",
    "stock_cap": "350 W",
    "version_why": "the newer of the two versions measured (0.27.1 on 2026-09-28, 0.30.0 on 2026-10-01)"
   },
   {
    "caps": [
     "350 W"
    ],
    "card": "rtx-3090",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "350 W"
      ],
      "complete": 4,
      "how": "anchor",
      "max": "146.308",
      "max_shown": "146.3",
      "min": "135.249",
      "min_shown": "135.2",
      "n": 4,
      "r8_left_out": 6,
      "ref": "r15",
      "refs": [
       "r16",
       "r15",
       "r17",
       "r18"
      ],
      "shown": "136.3"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": "350 W",
    "notes": [],
    "slug": "rtx-3090/gemma4-26b/ollama",
    "stock_cap": "350 W",
    "version_why": "the version is stated; the other rows' exhibit does not state one"
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-3090",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "mistral-nemo",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 5,
      "how": "anchor",
      "max": "88.203",
      "max_shown": "88.2",
      "min": "60.41",
      "min_shown": "60.4",
      "n": 10,
      "r8_left_out": 0,
      "ref": "r19",
      "refs": [
       "r20",
       "r21",
       "r22",
       "r23",
       "r24",
       "r19",
       "r25",
       "r26",
       "r27",
       "r28"
      ],
      "shown": "74.9"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.54822",
      "max_shown": "0.55",
      "min": "0.4818",
      "min_shown": "0.48",
      "n": 5,
      "r8_left_out": 0,
      "ref": "r29",
      "refs": [
       "r30",
       "r29",
       "r31",
       "r32",
       "r33"
      ],
      "shown": "0.51"
     }
    },
    "label": "mistral-nemo:12b-instruct-2407",
    "model": "mistral-nemo:12b-instruct-2407",
    "narrowed_to": null,
    "notes": [],
    "slug": "rtx-3090/mistral-nemo-12b-instruct-2407/ollama",
    "stock_cap": "350 W",
    "version_why": null
   },
   {
    "caps": [
     "350 W"
    ],
    "card": "rtx-3090",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "350 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "46.88",
      "max_shown": "46.9",
      "min": "46.88",
      "min_shown": "46.9",
      "n": 1,
      "r8_left_out": 3,
      "ref": "r34",
      "refs": [
       "r34"
      ],
      "shown": "46.9"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": "350 W",
    "notes": [],
    "slug": "rtx-3090/mistral-small3.2-24b/ollama",
    "stock_cap": "350 W",
    "version_why": "the version is stated; the other rows' exhibit does not state one"
   },
   {
    "caps": [
     "450 W"
    ],
    "card": "rtx-3090-ti",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "450 W"
      ],
      "complete": 2,
      "how": "anchor",
      "max": "147.56",
      "max_shown": "147.6",
      "min": "147.364",
      "min_shown": "147.4",
      "n": 2,
      "r8_left_out": 8,
      "ref": "r35",
      "refs": [
       "r35",
       "r36"
      ],
      "shown": "147.4"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": "450 W",
    "notes": [],
    "slug": "rtx-3090-ti/gemma4-26b/ollama",
    "stock_cap": "450 W",
    "version_why": null
   },
   {
    "caps": [
     "450 W"
    ],
    "card": "rtx-3090-ti",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "450 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "57.396",
      "max_shown": "57.4",
      "min": "57.396",
      "min_shown": "57.4",
      "n": 1,
      "r8_left_out": 3,
      "ref": "r37",
      "refs": [
       "r37"
      ],
      "shown": "57.4"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": "450 W",
    "notes": [],
    "slug": "rtx-3090-ti/mistral-small3.2-24b/ollama",
    "stock_cap": "450 W",
    "version_why": null
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-3090-ti",
    "engine": "llama.cpp",
    "engine_version": "b11514",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "22.657",
      "max_shown": "22.7",
      "min": "22.657",
      "min_shown": "22.7",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r38",
      "refs": [
       "r38"
      ],
      "shown": "22.7"
     },
     "prefill": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "255.2",
      "max_shown": "255",
      "min": "255.2",
      "min_shown": "255",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r39",
      "refs": [
       "r39"
      ],
      "shown": "255"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · llama.cpp",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "This run came last in its window, with the model's weights already in the computer's file cache (0 major page faults), per its exhibit.",
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-3090-ti/qwen3.8-flash-next/llama.cpp",
    "stock_cap": "450 W",
    "version_why": "the current build, 452 commits after 3cf03257, per its exhibit"
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-3090-ti",
    "engine": "the Strata engine",
    "engine_version": "0.1.39",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "98.291",
      "max_shown": "98.3",
      "min": "98.291",
      "min_shown": "98.3",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r40",
      "refs": [
       "r40"
      ],
      "shown": "98.3"
     },
     "prefill": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "2143.1",
      "max_shown": "2,143",
      "min": "2143.1",
      "min_shown": "2,143",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r41",
      "refs": [
       "r41"
      ],
      "shown": "2143"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · the Strata engine",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-3090-ti/qwen3.8-flash-next/the-strata-engine",
    "stock_cap": "450 W",
    "version_why": null
   },
   {
    "caps": [
     "95 W"
    ],
    "card": "rtx-5090-laptop-gpu",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "95 W"
      ],
      "complete": 4,
      "how": "anchor",
      "max": "121.379",
      "max_shown": "121.4",
      "min": "116.125",
      "min_shown": "116.1",
      "n": 4,
      "r8_left_out": 5,
      "ref": "r42",
      "refs": [
       "r43",
       "r42",
       "r44",
       "r45"
      ],
      "shown": "118.4"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "95 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.43596",
      "max_shown": "0.44",
      "min": "0.32627",
      "min_shown": "0.33",
      "n": 4,
      "r8_left_out": 5,
      "ref": "r46",
      "refs": [
       "r47",
       "r46",
       "r48",
       "r49"
      ],
      "shown": "0.35"
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": "95 W",
    "notes": [],
    "slug": "rtx-5090-laptop-gpu/gemma4-26b/ollama",
    "stock_cap": "95 W",
    "version_why": null
   },
   {
    "caps": [
     "95 W"
    ],
    "card": "rtx-5090-laptop-gpu",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "minicpm-v4.5",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "95 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "86.808",
      "max_shown": "86.8",
      "min": "86.808",
      "min_shown": "86.8",
      "n": 1,
      "r8_left_out": 1,
      "ref": "r50",
      "refs": [
       "r50"
      ],
      "shown": "86.8"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "95 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.14986",
      "max_shown": "0.15",
      "min": "0.14986",
      "min_shown": "0.15",
      "n": 1,
      "r8_left_out": 1,
      "ref": "r51",
      "refs": [
       "r51"
      ],
      "shown": "0.15"
     }
    },
    "label": "minicpm-v4.5",
    "model": "minicpm-v4.5",
    "narrowed_to": "95 W",
    "notes": [],
    "slug": "rtx-5090-laptop-gpu/minicpm-v4.5/ollama",
    "stock_cap": "95 W",
    "version_why": null
   },
   {
    "caps": [
     "95 W"
    ],
    "card": "rtx-5090-laptop-gpu",
    "engine": "Ollama",
    "engine_version": "0.32.13",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "95 W"
      ],
      "complete": 2,
      "how": "anchor",
      "max": "28.851",
      "max_shown": "28.9",
      "min": "28.839",
      "min_shown": "28.8",
      "n": 2,
      "r8_left_out": 3,
      "ref": "r52",
      "refs": [
       "r52",
       "r53"
      ],
      "shown": "28.8"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [
       "95 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "0.21437",
      "max_shown": "0.21",
      "min": "0.21094",
      "min_shown": "0.21",
      "n": 2,
      "r8_left_out": 3,
      "ref": "r54",
      "refs": [
       "r54",
       "r55"
      ],
      "shown": "0.21"
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": "95 W",
    "notes": [],
    "slug": "rtx-5090-laptop-gpu/mistral-small3.2-24b/ollama",
    "stock_cap": "95 W",
    "version_why": null
   },
   {
    "caps": [
     "135–150 W"
    ],
    "card": "rtx-5090-laptop-gpu",
    "engine": "llama.cpp",
    "engine_version": "3cf03257",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "135–150 W"
      ],
      "complete": 2,
      "how": "anchor",
      "max": "32.5",
      "max_shown": "32.5",
      "min": "22.5",
      "min_shown": "22.5",
      "n": 3,
      "r8_left_out": 0,
      "ref": "r56",
      "refs": [
       "r56",
       "r57",
       "r58"
      ],
      "shown": "22.5"
     },
     "prefill": {
      "caps": [
       "135–150 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "533.7",
      "max_shown": "534",
      "min": "211.7",
      "min_shown": "212",
      "n": 2,
      "r8_left_out": 0,
      "ref": "r59",
      "refs": [
       "r60",
       "r59"
      ],
      "shown": "534"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · llama.cpp",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-5090-laptop-gpu/qwen3.8-flash-next/llama.cpp",
    "stock_cap": "95 W",
    "version_why": null
   },
   {
    "caps": [
     "135–150 W"
    ],
    "card": "rtx-5090-laptop-gpu",
    "engine": "the Strata engine",
    "engine_version": "0.1.25",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "135–150 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "85.6",
      "max_shown": "85.6",
      "min": "85.6",
      "min_shown": "85.6",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r61",
      "refs": [
       "r61"
      ],
      "shown": "85.6"
     },
     "prefill": {
      "caps": [
       "135–150 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "2147.5",
      "max_shown": "2,148",
      "min": "2147.5",
      "min_shown": "2,148",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r62",
      "refs": [
       "r62"
      ],
      "shown": "2148"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · the Strata engine",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-5090-laptop-gpu/qwen3.8-flash-next/the-strata-engine",
    "stock_cap": "95 W",
    "version_why": null
   },
   {
    "caps": [
     "420 W"
    ],
    "card": "rtx-pro-6000-blackwell-workstation-edition",
    "engine": "Ollama",
    "engine_version": "not stated",
    "family": "gemma4",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "420 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "205.85",
      "max_shown": "205.9",
      "min": "205.85",
      "min_shown": "205.9",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r63",
      "refs": [
       "r63"
      ],
      "shown": "205.9"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "gemma4:26b",
    "model": "gemma4:26b",
    "narrowed_to": null,
    "notes": [],
    "slug": "rtx-pro-6000-blackwell-workstation-edition/gemma4-26b/ollama",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "420 W"
    ],
    "card": "rtx-pro-6000-blackwell-workstation-edition",
    "engine": "Ollama",
    "engine_version": "not stated",
    "family": "mistral-small",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "420 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "92.77",
      "max_shown": "92.8",
      "min": "92.77",
      "min_shown": "92.8",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r64",
      "refs": [
       "r64"
      ],
      "shown": "92.8"
     },
     "prefill": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "mistral-small3.2:24b",
    "model": "mistral-small3.2:24b",
    "narrowed_to": null,
    "notes": [],
    "slug": "rtx-pro-6000-blackwell-workstation-edition/mistral-small3.2-24b/ollama",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-pro-6000-blackwell-workstation-edition",
    "engine": "llama.cpp",
    "engine_version": "3cf03257",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "102.889",
      "max_shown": "102.9",
      "min": "102.889",
      "min_shown": "102.9",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r65",
      "refs": [
       "r65"
      ],
      "shown": "102.9"
     },
     "prefill": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "1619.4",
      "max_shown": "1,619",
      "min": "1619.4",
      "min_shown": "1,619",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r66",
      "refs": [
       "r66"
      ],
      "shown": "1619"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · llama.cpp",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-pro-6000-blackwell-workstation-edition/qwen3.8-flash-next/llama.cpp",
    "stock_cap": null,
    "version_why": null
   },
   {
    "caps": [
     "300 W"
    ],
    "card": "rtx-pro-6000-blackwell-workstation-edition",
    "engine": "the Strata engine",
    "engine_version": "0.1.39",
    "family": "qwen3.8",
    "family_stand_in": false,
    "inputs": {
     "decode": {
      "caps": [
       "300 W"
      ],
      "complete": 1,
      "how": "anchor",
      "max": "205.263",
      "max_shown": "205.3",
      "min": "205.263",
      "min_shown": "205.3",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r67",
      "refs": [
       "r67"
      ],
      "shown": "205.3"
     },
     "prefill": {
      "caps": [
       "300 W"
      ],
      "complete": null,
      "how": "same run",
      "max": "3920.0",
      "max_shown": "3,920",
      "min": "3920.0",
      "min_shown": "3,920",
      "n": 1,
      "r8_left_out": 0,
      "ref": "r68",
      "refs": [
       "r68"
      ],
      "shown": "3920"
     },
     "ttft": {
      "caps": [],
      "complete": null,
      "how": null,
      "max": null,
      "max_shown": null,
      "min": null,
      "min_shown": null,
      "n": 0,
      "r8_left_out": 0,
      "ref": null,
      "refs": [],
      "shown": null
     }
    },
    "label": "Qwen3.8-Flash-Next · the Strata engine",
    "model": "Qwen3.8-Flash-Next",
    "narrowed_to": null,
    "notes": [
     "We measured this pair's wait only on a 4,038-token question, so the wait here is worked out from its reading speed and the question's length in the settings below: reading only."
    ],
    "slug": "rtx-pro-6000-blackwell-workstation-edition/qwen3.8-flash-next/the-strata-engine",
    "stock_cap": null,
    "version_why": null
   }
  ]
 },
 "prompt_tokens": 519,
 "rows": {
  "r1": {
   "also": [],
   "card": "rtx-3080",
   "context": "not stated",
   "engine": "not stated",
   "exhibit": "a-short-history-of-mistral",
   "figure": "34.40 tokens a second",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second, split with system memory",
   "limit": "not stated",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "what-we-measured-on-two-cards-in-one-box",
   "settings": "not stated",
   "version": "not stated",
   "weights": "not stated"
  },
  "r10": {
   "also": [
    {
     "exhibit": "two-used-3080s-priced",
     "section": "power-heat-and-the-fifty-watts"
    }
   ],
   "card": "two-rtx-3080",
   "context": "49,152",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "44.7",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "300 W each, 600 W for both cards",
   "measured": "2026-09-17",
   "model": "mistral-small3.2:24b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "how-fast-does-each-one-write",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r11": {
   "also": [],
   "card": "rtx-3090",
   "context": "16 slots of 5,120",
   "engine": "Ollama",
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "85.007",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode per request, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-01 (UTC)",
   "model": "Gemma 4 12B",
   "output_tokens": 256,
   "prompt": "536 tokens",
   "section": "requests-at-once",
   "settings": "256 tokens out · `num_batch: 512`",
   "version": "0.34.4",
   "weights": "gemma4:12b-it-qat, a 4-bit GGUF"
  },
  "r12": {
   "also": [],
   "card": "rtx-3090",
   "context": "16 slots of 5,120",
   "engine": "Ollama",
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "441.866",
   "input": "ttft",
   "kv_cache": "not stated",
   "label": "the wait for the first token, 95th percentile, ms",
   "limit": "300 W",
   "measured": "2026-10-01 (UTC)",
   "model": "Gemma 4 12B",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "requests-at-once",
   "settings": "not stated",
   "version": "0.34.4",
   "weights": "gemma4:12b-it-qat, a 4-bit GGUF"
  },
  "r13": {
   "also": [],
   "card": "rtx-3090",
   "context": "5,120",
   "engine": "vLLM",
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "82.693",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode per request, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-01 (UTC)",
   "model": "Gemma 4 12B",
   "output_tokens": 256,
   "prompt": "536 tokens",
   "section": "requests-at-once",
   "settings": "256 tokens out · 512 tokens per step",
   "version": "0.30.0",
   "weights": "gemma-4-12B-it-qat-w4a16-ct"
  },
  "r14": {
   "also": [],
   "card": "rtx-3090",
   "context": "5,120",
   "engine": "vLLM",
   "exhibit": "ollama-and-vllm-on-one-rtx-3090",
   "figure": "212.999",
   "input": "ttft",
   "kv_cache": "not stated",
   "label": "the wait for the first token, 95th percentile, ms",
   "limit": "300 W",
   "measured": "2026-10-01 (UTC)",
   "model": "Gemma 4 12B",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "requests-at-once",
   "settings": "not stated",
   "version": "0.30.0",
   "weights": "gemma-4-12B-it-qat-w4a16-ct"
  },
  "r15": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "136.274",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode summed across conversations, tokens a second",
   "limit": "350 W",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "three-more-arms",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r16": {
   "also": [],
   "card": "rtx-3090",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "135.249, 19,415 MiB",
   "input": "decode",
   "kv_cache": "q4_0",
   "label": "decode at the full window, and the memory it used",
   "limit": "350 W",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-fits-and-what-the-cache-costs",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r17": {
   "also": [
    {
     "exhibit": "one-3090-against-one-3090-ti",
     "section": "what-fits-and-what-the-cache-costs"
    }
   ],
   "card": "rtx-3090",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "136.618",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "350 W",
   "measured": "2026-09-18, 00:22 to 00:54 UTC",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r18": {
   "also": [],
   "card": "rtx-3090",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "146.308 tok/s, 20,995 MiB",
   "input": "decode",
   "kv_cache": "f16",
   "label": "decode at the full window, and the memory it used",
   "limit": "350 W",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-fits-and-what-the-cache-costs",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r19": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "74.894 (0.398)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q3_K_M"
  },
  "r2": {
   "also": [],
   "card": "rtx-3080",
   "context": "not stated",
   "engine": "not stated",
   "exhibit": "a-short-history-of-mistral",
   "figure": "5.13",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second, split with system memory",
   "limit": "not stated",
   "measured": "2026-09-10",
   "model": "mistral-small3.2:24b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "what-we-measured-on-two-cards-in-one-box",
   "settings": "not stated",
   "version": "not stated",
   "weights": "not stated"
  },
  "r20": {
   "also": [],
   "card": "rtx-3090",
   "context": "32,768",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "60.41",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q8_0"
  },
  "r21": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "60.545 (0.033)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q8_0"
  },
  "r22": {
   "also": [],
   "card": "rtx-3090",
   "context": "32,768",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "64.83",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q6_K"
  },
  "r23": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "65.724 (0.441)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q6_K"
  },
  "r24": {
   "also": [],
   "card": "rtx-3090",
   "context": "32,768",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "74.135",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q3_K_M"
  },
  "r25": {
   "also": [],
   "card": "rtx-3090",
   "context": "32,768",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "78.158",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q5_K_M"
  },
  "r26": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "78.942 (0.068)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q5_K_M"
  },
  "r27": {
   "also": [],
   "card": "rtx-3090",
   "context": "32,768",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "87.234",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r28": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "88.203 (0.252)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": 256,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r29": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "510.56",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": null,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q3_K_M"
  },
  "r3": {
   "also": [
    {
     "exhibit": "two-used-3080s-priced",
     "section": "what-the-cache-type-costs"
    },
    {
     "exhibit": "two-used-3080s-priced",
     "section": "power-heat-and-the-fifty-watts"
    }
   ],
   "card": "two-rtx-3080",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "115.4",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "300 W each, 600 W for both cards",
   "measured": "2026-09-17",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "how-fast-does-each-one-write",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r30": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "481.8",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": null,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r31": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "516.45",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": null,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q5_K_M"
  },
  "r32": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "518.25",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": null,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q6_K"
  },
  "r33": {
   "also": [],
   "card": "rtx-3090",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "the-quantization-ladder",
   "figure": "548.22",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "300 W",
   "measured": "2026-10-15 between 05:34 and 11:01 UTC",
   "model": "mistral-nemo:12b-instruct-2407",
   "output_tokens": null,
   "prompt": "519 tokens",
   "section": "speed-and-memory",
   "settings": "asks for 256 tokens",
   "version": "0.32.13",
   "weights": "Q8_0"
  },
  "r34": {
   "also": [
    {
     "exhibit": "one-3090-against-one-3090-ti",
     "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
    }
   ],
   "card": "rtx-3090",
   "context": "65,536",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "46.880",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "350 W",
   "measured": "2026-09-18, 00:22 to 00:54 UTC",
   "model": "mistral-small3.2:24b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r35": {
   "also": [],
   "card": "rtx-3090-ti",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "147.364",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode summed across conversations, tokens a second",
   "limit": "450 W",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "three-more-arms",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r36": {
   "also": [
    {
     "exhibit": "one-3090-against-one-3090-ti",
     "section": "the-two-models-want-different-caps"
    }
   ],
   "card": "rtx-3090-ti",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "147.560",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "450 W",
   "measured": "2026-09-17 (14:54 to 23:46 UTC)",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r37": {
   "also": [
    {
     "exhibit": "one-3090-against-one-3090-ti",
     "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
    },
    {
     "exhibit": "one-3090-against-one-3090-ti",
     "section": "the-two-models-want-different-caps"
    }
   ],
   "card": "rtx-3090-ti",
   "context": "65,536",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "57.396",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "450 W",
   "measured": "2026-09-17 (14:54 to 23:46 UTC)",
   "model": "mistral-small3.2:24b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r38": {
   "also": [],
   "card": "rtx-3090-ti",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "22.657",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-09 (09:32Z to 10:51Z, UTC)",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window",
   "settings": "no draft head · 8 of 64 threads",
   "version": "b11514",
   "weights": "IQ3_S"
  },
  "r39": {
   "also": [],
   "card": "rtx-3090-ti",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "255.2",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "300 W",
   "measured": "2026-10-09 (09:32Z to 10:51Z, UTC)",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window",
   "settings": "no draft head · 8 of 64 threads",
   "version": "b11514",
   "weights": "IQ3_S"
  },
  "r4": {
   "also": [],
   "card": "two-rtx-3080",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "113.8 tok/s, 18,815 MiB",
   "input": "decode",
   "kv_cache": "q4_0",
   "label": "decode at the full window, and the memory it used",
   "limit": "not stated",
   "measured": "2026-09-17",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-the-cache-type-costs",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r40": {
   "also": [],
   "card": "rtx-3090-ti",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "98.291",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-09 (09:32Z to 10:51Z, UTC)",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window",
   "settings": "8,144 of 24,576 experts on the card",
   "version": "0.1.39",
   "weights": "IQ3_S"
  },
  "r41": {
   "also": [],
   "card": "rtx-3090-ti",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,143.1",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "300 W",
   "measured": "2026-10-09 (09:32Z to 10:51Z, UTC)",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-24-gb-desktop-card-both-engines-in-one-window",
   "settings": "8,144 of 24,576 experts on the card",
   "version": "0.1.39",
   "weights": "IQ3_S"
  },
  "r42": {
   "also": [
    {
     "exhibit": "a-laptop-asked-the-desktops-questions",
     "section": "the-tables"
    }
   ],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "118.368",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "summed tokens/s",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r43": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "116.125 (115.846–116.716)",
   "input": "decode",
   "kv_cache": "q4_0",
   "label": "tokens/s (spread of three)",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r44": {
   "also": [
    {
     "exhibit": "a-laptop-asked-the-desktops-questions",
     "section": "the-tables"
    }
   ],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "118.424",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "tokens/s",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r45": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "121.379 (120.822–123.017)",
   "input": "decode",
   "kv_cache": "f16",
   "label": "tokens/s (spread of three)",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r46": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "345.52",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r47": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "326.27",
   "input": "ttft",
   "kv_cache": "q4_0",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r48": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "372.49",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r49": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "435.96",
   "input": "ttft",
   "kv_cache": "f16",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "gemma4:26b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r5": {
   "also": [
    {
     "exhibit": "two-used-3080s-priced",
     "section": "power-heat-and-the-fifty-watts"
    }
   ],
   "card": "two-rtx-3080",
   "context": "131,072",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "114.8",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "250 W each, 500 W for both cards",
   "measured": "2026-09-17",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "how-fast-does-each-one-write",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r50": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "86.808 (85.795–87.364)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "tokens/s (spread of three)",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "minicpm-v4.5",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "its text path only",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r51": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "149.86",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "minicpm-v4.5",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "its text path only",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r52": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "65,536",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "28.839",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "tokens/s",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "mistral-small3.2:24b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r53": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "98,304",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "28.851 (28.721–28.993)",
   "input": "decode",
   "kv_cache": "8-bit",
   "label": "tokens/s (spread of three)",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "mistral-small3.2:24b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "every layer forced onto the card",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r54": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "65,536",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "210.94",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "mistral-small3.2:24b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r55": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "98,304",
   "engine": "Ollama",
   "exhibit": "a-laptop-asked-the-desktops-questions",
   "figure": "214.37",
   "input": "ttft",
   "kv_cache": "8-bit",
   "label": "first token, ms",
   "limit": "95 W",
   "measured": "18:43 to 19:34 UTC on 2026-09-21",
   "model": "mistral-small3.2:24b",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "the-tables",
   "settings": "every layer forced onto the card",
   "version": "0.32.13",
   "weights": "not stated"
  },
  "r56": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "22.5",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "135–150 W",
   "measured": "2026-10-05T14:43:47Z → 14:59:47Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file",
   "settings": "a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r57": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "24.4",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "135–150 W",
   "measured": "2026-10-02T23:03:15Z → 23:40:55Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file",
   "settings": "a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r58": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "32.5",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "135–150 W",
   "measured": "2026-10-03T06:07:35Z → 06:31:09Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-q2-0-pair-the-laptop",
   "settings": "8 threads · 24 of 48 layers of experts in system memory",
   "version": "3cf03257",
   "weights": "Q2_0"
  },
  "r59": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "533.7",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "135–150 W",
   "measured": "2026-10-05T14:43:47Z → 14:59:47Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-first-word-the-laptop-the-iq3-s-file",
   "settings": "a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r6": {
   "also": [],
   "card": "two-rtx-3080",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "124.0",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode summed across conversations, tokens a second",
   "limit": "300 W each, 600 W for both cards",
   "measured": "2026-09-17",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "the-full-concurrency-ladder",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r60": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "211.7",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "135–150 W",
   "measured": "2026-10-02T23:03:15Z → 23:40:55Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-first-word-the-laptop-the-iq3-s-file",
   "settings": "a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r61": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "85.6",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "135–150 W",
   "measured": "2026-10-03T00:46:41Z → 00:50:45Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "writing-the-answer-the-laptop-the-iq3-s-file",
   "settings": "reading 8,192 tokens at a time · about a third of the experts on the card",
   "version": "0.1.25",
   "weights": "IQ3_S"
  },
  "r62": {
   "also": [],
   "card": "rtx-5090-laptop-gpu",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "2,147.5",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "135–150 W",
   "measured": "2026-10-03T00:46:41Z → 00:50:45Z",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-first-word-the-laptop-the-iq3-s-file",
   "settings": "reading 8,192 tokens at a time · about a third of the experts on the card",
   "version": "0.1.25",
   "weights": "IQ3_S"
  },
  "r63": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-short-history-of-mistral",
   "figure": "205.85 tokens a second",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "420 W",
   "measured": "the small hours of 2026-09-16 (UTC)",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-we-measured-on-two-cards-in-one-box",
   "settings": "256 tokens out",
   "version": "not stated",
   "weights": "not stated"
  },
  "r64": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "Ollama",
   "exhibit": "a-short-history-of-mistral",
   "figure": "92.77 tokens a second",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "420 W",
   "measured": "the small hours of 2026-09-16 (UTC)",
   "model": "mistral-small3.2:24b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "what-we-measured-on-two-cards-in-one-box",
   "settings": "256 tokens out",
   "version": "not stated",
   "weights": "Q4_K_M"
  },
  "r65": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "102.889",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-06",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-desktop-card-one-engine-a-night",
   "settings": "every layer on the card · 8 threads · a 512-token micro-batch",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r66": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "llama.cpp",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "1,619.4",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "300 W",
   "measured": "2026-10-06",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-desktop-card-one-engine-a-night",
   "settings": "every layer on the card · 8 threads · a 512-token micro-batch",
   "version": "3cf03257",
   "weights": "IQ3_S"
  },
  "r67": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "205.263",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "writing speed, tokens a second",
   "limit": "300 W",
   "measured": "2026-10-07",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-desktop-card-one-engine-a-night",
   "settings": "all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
   "version": "0.1.39",
   "weights": "IQ3_S"
  },
  "r68": {
   "also": [],
   "card": "rtx-pro-6000-blackwell-workstation-edition",
   "context": "not stated",
   "engine": "the Strata engine",
   "exhibit": "flash-next-why-the-first-word-comes-sooner",
   "figure": "3,920.0",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt rate, tokens a second, the engine's clock",
   "limit": "300 W",
   "measured": "2026-10-07",
   "model": "Qwen3.8-Flash-Next",
   "output_tokens": null,
   "prompt": "4K (4,038 tokens)",
   "section": "the-desktop-card-one-engine-a-night",
   "settings": "all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
   "version": "0.1.39",
   "weights": "IQ3_S"
  },
  "r69": {
   "also": [],
   "card": "rtx-3090",
   "context": "131,072, with about 98,000 tokens in it",
   "engine": "Ollama",
   "exhibit": "one-3090-against-one-3090-ti",
   "figure": "2,270.547",
   "input": "prefill",
   "kv_cache": "not stated",
   "label": "prompt reading, tokens a second",
   "limit": "350 W",
   "measured": "not stated",
   "model": "not stated",
   "output_tokens": null,
   "prompt": "not stated",
   "section": "three-more-arms",
   "settings": "not stated",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r7": {
   "also": [],
   "card": "two-rtx-3080",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "124.4",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode summed across conversations, tokens a second",
   "limit": "250 W each, 500 W for both cards",
   "measured": "2026-09-17",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "the-full-concurrency-ladder",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r8": {
   "also": [],
   "card": "two-rtx-3080",
   "context": "4,096",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "124.7",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode out of the box once the layer count is named",
   "limit": "not stated",
   "measured": "not stated",
   "model": "gemma4:26b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "at-a-glance",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  },
  "r9": {
   "also": [
    {
     "exhibit": "two-used-3080s-priced",
     "section": "power-heat-and-the-fifty-watts"
    }
   ],
   "card": "two-rtx-3080",
   "context": "49,152",
   "engine": "Ollama",
   "exhibit": "two-used-3080s-priced",
   "figure": "44.6",
   "input": "decode",
   "kv_cache": "not stated",
   "label": "decode, tokens a second",
   "limit": "250 W each, 500 W for both cards",
   "measured": "2026-09-17",
   "model": "mistral-small3.2:24b",
   "output_tokens": 256,
   "prompt": "not stated",
   "section": "how-fast-does-each-one-write",
   "settings": "256 tokens out",
   "version": "0.32.13",
   "weights": "Q4_K_M"
  }
 },
 "schema": "speed-picker/1"
}