{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "flash-next-on-three-cards",
 "title": "Qwen3.8-Flash-Next on an RTX 5090 laptop, an RTX PRO 6000 Blackwell and an RTX 3090 Ti: the Strata engine and llama.cpp, the numbers",
 "dek": "Qwen3.8-Flash-Next, a model too large to hold on one 24 GB card, run one request at a time through llama.cpp (at commit 3cf03257, without the draft head, which the file we benched does not carry) and the Strata engine (github.com/Niko1221/Strata, an unrelated project) on an RTX 5090 laptop card (24 GB, in a laptop with 64 GB of memory, no settable cap), an RTX PRO 6000 Blackwell (96 GB, at a 300 W cap, one engine a night) and an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB, at a 300 W cap, both engines in one window beside live work on the computer's other card): tokens a second, the wait for the first word and board watts, side by side in tables and charts, each figure with the conditions it was measured under and printed first on the full bench, with one paragraph per card, the findings, which card and engine for which case, and what it does not say.",
 "published": "2026-10-12",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/flash-next-on-three-cards/",
  "md_url": "https://research.strata2signal.com/flash-next-on-three-cards/index.md",
  "md_sha": "5e4d3b0745a6797faf257fb2d3e682f73a6bcd091521d4015088fce525f76e65",
  "html_sha": "df907236db5d7c6a9f1e405f141e82059076e8185d7c819dcce2ab405bb5db23",
  "anchors_sha": "e0587e11be564575e14419497405316dea6cbca0dc6076697259c0fc18fef09f"
 },
 "short": {
  "paragraph": "Qwen3.8-Flash-Next is a large AI model, released as a set of files. A program that loads those files and answers requests is an engine. llama.cpp is a widely used open-source engine for such files; the Strata engine (github.com/Niko1221/Strata) is a newer engine built on llama.cpp that runs this one model. Despite the similar name, Strata (github.com/Niko1221/Strata) is an unrelated project; strata2signal has no affiliation with it. We ran both engines on three graphics cards, one request at a time, with the same prompts, and timed them: writing speed in tokens a second (a token is a word or a piece of one, the unit an engine reads and writes), the wait for the first word in seconds, and the card's own board watts, each figure beside the conditions it was measured under, and no ratio drawn between cards.",
  "from_dek": false,
  "counts": {
   "words": 11016,
   "minutes": 50,
   "tables": 5,
   "kit": false
  },
  "bullets": [
   {
    "text": "the Strata engine 0.1.25 on a 5090 laptop waited 1.9 s",
    "figure": "1.9 s",
    "cite": "the-first-word",
    "quote": "RTX 5090 Laptop: llama.cpp 19.1 s; The Strata engine 1.9 s.",
    "span": [
     30211,
     30217
    ]
   },
   {
    "text": "the Strata engine 0.1.25 on a 5090 laptop drew 126.3 W",
    "figure": "126.3 W",
    "cite": "what-each-card-drew",
    "quote": "RTX 5090 Laptop: llama.cpp 61.1 W; The Strata engine 126.3 W.",
    "span": [
     43826,
     43834
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-three-cards-and-how-they-were-run",
   "heading": "The three cards, and how they were run",
   "level": 2,
   "span": [
    8490,
    15205
   ],
   "chunks": [
    [
     8490,
     15205
    ]
   ],
   "chars": 6715,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "writing-the-answer",
   "heading": "Writing the answer",
   "level": 2,
   "span": [
    15205,
    25038
   ],
   "chunks": [
    [
     15205,
     25038
    ]
   ],
   "chars": 9833,
   "digest": "this section presents token-per-second results for llama.cpp and the Strata engine on three different cards. it compares performance at 4K and 32K prompt lengths, noting the impact of the draft head and specific engine builds.",
   "digest_skipped": null
  },
  {
   "id": "the-first-word",
   "heading": "The first word",
   "level": 2,
   "span": [
    25038,
    39382
   ],
   "chunks": [
    [
     25038,
     39382
    ]
   ],
   "chars": 14344,
   "digest": "this section reports the time in seconds from sending a prompt to receiving the first token. it covers 4K and 32K prompt lengths for the laptop, an RTX PRO 6000, and an RTX 3090 Ti using both engines.",
   "digest_skipped": null
  },
  {
   "id": "what-each-card-drew",
   "heading": "What each card drew",
   "level": 2,
   "span": [
    39382,
    46373
   ],
   "chunks": [
    [
     39382,
     46373
    ]
   ],
   "chars": 6991,
   "digest": "this section records the board power in watts for each card and engine during 4K and 32K requests. it notes that the desktop cards were under a 300 W cap and specifies the number of samples per request.",
   "digest_skipped": null
  },
  {
   "id": "each-card-in-one-paragraph",
   "heading": "Each card, in one paragraph",
   "level": 2,
   "span": [
    46373,
    53148
   ],
   "chunks": [
    [
     46373,
     53148
    ]
   ],
   "chars": 6775,
   "digest": "this section provides a detailed summary for each card. it describes how the laptop, the 96 GB card, and the 24 GB desktop card handled model experts in memory, power limits, and engine performance characteristics.",
   "digest_skipped": null
  },
  {
   "id": "what-we-found",
   "heading": "What we found",
   "level": 2,
   "span": [
    53148,
    60100
   ],
   "chunks": [
    [
     53148,
     60100
    ]
   ],
   "chars": 6952,
   "digest": "this section summarizes key findings, including the Strata engine's speed advantages on 24 GB and 96 GB cards, the role of the draft head, and the fact that system memory is the gate for the Strata engine.",
   "digest_skipped": null
  },
  {
   "id": "which-card-and-engine-for-which-case",
   "heading": "Which card and engine for which case",
   "level": 2,
   "span": [
    60100,
    65909
   ],
   "chunks": [
    [
     60100,
     65909
    ]
   ],
   "chars": 5809,
   "digest": "this section offers recommendations based on specific use cases. it suggests engines and hardware configurations for single users, 96 GB card users, 24 GB desktop card users, and scenarios involving parallel calls or specific memory limits.",
   "digest_skipped": null
  },
  {
   "id": "what-it-does-not-say",
   "heading": "What it does not say",
   "level": 2,
   "span": [
    65909,
    67777
   ],
   "chunks": [
    [
     65909,
     67777
    ]
   ],
   "chars": 1868,
   "digest": "this section clarifies the limitations of the data, stating that no ratios are compared between cards, it does not cover wall power, and the llama.cpp runs were not at their tuned best.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-where-the-files-are",
   "heading": "Who ran this, and where the files are",
   "level": 2,
   "span": [
    67777,
    68804
   ],
   "chunks": [
    [
     67777,
     68804
    ]
   ],
   "chars": 1027,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-e19d84ad",
   "q": "how do the engines compare in token generation speed?",
   "a": "The Strata engine generally outperformed llama.cpp in tokens per second. For example, on the RTX 5090 Laptop at 4K, the Strata engine 0.1.25 achieved 85.6 tok/s compared to llama.cpp's 24.4 tok/s. On the RTX PRO 6000, the Strata engine 0.1.39 reached 205.263 tok/s while llama.cpp reached 102.889 tok/s.",
   "cites": [
    "writing-the-answer"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-b216162d",
   "q": "what are the power consumption levels for the cards?",
   "a": "The RTX 5090 Laptop drew 61.1 W to 140.3 W depending on the engine and prompt length. The RTX PRO 6000 stayed near its 300 W cap, with readings like 297.8 W at 4K. The RTX 3090 Ti drew between 168.5 W and 293.7 W across the various tests.",
   "cites": [
    "what-each-card-drew"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "flash-next-why-the-first-word-comes-sooner",
   "why": "shares ground with § Writing the answer · § What this page is"
  },
  {
   "slug": "six-cards-the-same-questions",
   "why": "shares ground with § What each card drew · § What the boards actually drew"
  },
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § What each card drew · § What a power cap takes, and what each card does at its own limit"
  }
 ],
 "thanks": "## Who ran this, and where the files are {#who-ran-this-and-where-the-files-are}\n\nThe Strata engine's author (github.com/Niko1221/Strata; MIT); Georgi Gerganov and the ggml authors (llama.cpp, MIT); ISTA-DASLab (the GGUF files, the compressed copies of the weights in the file format llama.cpp reads); Qwen (the model and its draft head, under the Qwen Community License 1.0); NVIDIA (CUDA). None of them owed us anything. A small (human) team wrote the questions, pressed the keys and read the pages; a fleet of AI agents built the harness, ran the audits and drafted the prose. Every figure on this page is printed first on [Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench](https://research.strata2signal.com/flash-next-why-the-first-word-comes-sooner/), with its row, file and field; this page adds no figure the full bench does not print. Despite the similar name, Strata (github.com/Niko1221/Strata) is an unrelated project; strata2signal has no affiliation with it.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 8,
  "digests_kept": 7,
  "chips_written": 7,
  "chips_kept": 2,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 1,
   "figure": 0,
   "length": 3,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 1,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 1,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 1
  },
  "new_noun_tokens_checked": [
   "0.1.25",
   "0.1.39",
   "1.9",
   "102.889",
   "12.14",
   "126.3",
   "138.02",
   "140.3",
   "168.5",
   "19.1",
   "205.263",
   "24",
   "24.4",
   "293.7",
   "297.8",
   "3.5",
   "300",
   "3090",
   "32",
   "32K",
   "4.9",
   "46.84",
   "4K",
   "5090",
   "6000",
   "61.1",
   "64",
   "85.6",
   "96",
   "EVGA",
   "Edition",
   "FTW3",
   "GB",
   "GPU",
   "GeForce",
   "GiB",
   "Laptop",
   "PRO",
   "RTX",
   "Strata",
   "Ti",
   "Ultra",
   "W",
   "Workstation"
  ],
  "new_noun_tokens_withheld": 2,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "new_noun",
    "item": null,
    "item_chars": 4,
    "withheld": "a string naming something the article does not",
    "detail": ""
   },
   {
    "kind": "bullet",
    "reason": "bare_figure",
    "item": "the Strata engine 0.1.39 on an RTX PRO 6000 wrote 205.263",
    "detail": "the cited section writes this figure as '205.263 tok/s' and the bullet prints '205.263' alone"
   },
   {
    "kind": "chip",
    "reason": "redaction",
    "item": null,
    "item_chars": 349,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "length",
    "item": "what are the main performance advantages of the Strata engine?",
    "detail": "q not lowercase"
   },
   {
    "kind": "chip",
    "reason": "length",
    "item": "which engine is recommended for a single user with a 24 GB card?",
    "detail": "q not lowercase"
   },
   {
    "kind": "chip",
    "reason": "length",
    "item": "what limitations does the Strata engine have regarding memory?",
    "detail": "q not lowercase"
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 315,
    "text_sha": "fd91cac70bf7",
    "chip_id": "c-11560e2e",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-12",
    "why": "Opens with a blanket claim the section refuses on the 96 GB card at 4K (within unmeasured night-to-night variation, no ratio), and prints the RTX 3090 Ti's 32K pair without its exploratory label.",
    "detail": "removed from this pack by hand on 2026-10-12 (UTC), after the writer ran and after the gates had counted: Opens with a blanket claim the section refuses on the 96 GB card at 4K (within unmeasured night-to-night variation, no ratio), and prints the RTX 3090 Ti's 32K pair without its exploratory label. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-10-12T14:38:58Z"
}