{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "ollama-and-vllm-on-one-rtx-3090",
 "title": "Ollama and vLLM on one RTX 3090",
 "dek": "Ollama and vLLM both run an open language model on your own computer. This page puts both on one RTX 3090 (24 GB) with the same 4-bit Gemma 4 12B weights, from one request at a time to sixteen at once, and prints every reading beside the bet we wrote down before it.",
 "published": "2026-10-07",
 "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/ollama-and-vllm-on-one-rtx-3090/",
  "md_url": "https://research.strata2signal.com/ollama-and-vllm-on-one-rtx-3090/index.md",
  "md_sha": "631bb8c28625dd9e7ad47c65e4144173483bd5e3350df05c19b748282eb88435",
  "html_sha": "385f67d8c0d3a45cda2496d9dbe71b6c89c9786385655ec6d8505a3ea86e6756",
  "anchors_sha": "4c39bc5db0169190c71c78cd48b749d75c4b2fffaddf47b97118659457771f63"
 },
 "short": {
  "paragraph": "Ollama and vLLM are two ways to run an open model on your own computer; this page measures both on one RTX 3090 (24 GB) with the same 4-bit Gemma 4 12B weights. One request at a time, the two write at nearly the same speed once an answer has started: on the older pair we tested first, Ollama 0.32.13 wrote 85.2 tokens a second and vLLM 0.27.1 wrote 78.1 tokens a second, inside the 10 per cent band we had predicted, and the prediction held again on Ollama 0.34.4 and vLLM 0.30.0. Many at once is the job vLLM was built for, and it shows: at sixteen requests released together, vLLM 0.30.0 wrote 506.829 tokens a second across them and Ollama 0.34.4, set to serve sixteen at once, 142.327 tokens a second. vLLM's total first reached 1.2 times Ollama's at two requests at once, a ratio of 1.2077, by a margin smaller than the variation between repeated runs; we had predicted four at once, and had written down before any run that two or eight at once would settle nothing either way, so that prediction stands undecided. Two predictions about the wait before an answer starts were wrong: at one request Ollama's wait was more than twice vLLM's, and with sixteen long prompts at once neither server kept that wait inside the limit we had set for it, 1.5 s for vLLM or 4 s for Ollama set to serve sixteen at once. One prediction about the answers was wrong too: on 108 lines of dinner-party talk, each asking which of six guests said it, the two servers agreed on fewer of them than the 97 per cent we had predicted, so every speed row here carries the label \"answers differ at this level\". It is one card and one model, with requests released together, not people arriving over time.",
  "from_dek": false,
  "counts": {
   "words": 11758,
   "minutes": 53,
   "tables": 20,
   "kit": true
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "the-words-this-page-leans-on",
   "heading": "The words this page leans on",
   "level": 2,
   "span": [
    3778,
    5804
   ],
   "chunks": [
    [
     3778,
     5804
    ]
   ],
   "chars": 2026,
   "digest": "the article defines terms like reading and writing, window, slot, requests at once, posture, arm, prediction, verdict, and spread. it establishes how decode speed, total speed, and the spread of a level are measured and reported.",
   "digest_skipped": null
  },
  {
   "id": "two-names-one-card",
   "heading": "Two names, one card",
   "level": 2,
   "span": [
    5804,
    8935
   ],
   "chunks": [
    [
     5804,
     8935
    ]
   ],
   "chars": 3131,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "one-request-at-a-time",
   "heading": "One request at a time",
   "level": 2,
   "span": [
    8935,
    12099
   ],
   "chunks": [
    [
     8935,
     12099
    ]
   ],
   "chars": 3164,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "requests-at-once",
   "heading": "Requests at once",
   "level": 2,
   "span": [
    12099,
    24546
   ],
   "chunks": [
    [
     12099,
     24546
    ]
   ],
   "chars": 12447,
   "digest": "arm C measured total speed and each request's writing speed for one to sixteen requests at once. it also tested a long prompt and measured board energy in joules per 1,000 tokens across different levels.",
   "digest_skipped": null
  },
  {
   "id": "the-first-word",
   "heading": "The first word, and a shared prompt",
   "level": 2,
   "span": [
    24546,
    33356
   ],
   "chunks": [
    [
     24546,
     33356
    ]
   ],
   "chars": 8810,
   "digest": "the article examines first-token latency and prompt caching. it describes how Ollama's load step changed between versions and how both servers use caches to serve shared prompts differently.",
   "digest_skipped": null
  },
  {
   "id": "what-each-holds",
   "heading": "What each holds, and how fast it starts",
   "level": 2,
   "span": [
    33356,
    39478
   ],
   "chunks": [
    [
     33356,
     39478
    ]
   ],
   "chars": 6122,
   "digest": "the text details how much memory each server uses at rest and under load. it notes that vLLM reserves memory at start, while Ollama's reservation follows its slots and window.",
   "digest_skipped": null
  },
  {
   "id": "are-the-answers-the-same",
   "heading": "Are the answers the same?",
   "level": 2,
   "span": [
    39478,
    42410
   ],
   "chunks": [
    [
     39478,
     42410
    ]
   ],
   "chars": 2932,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-was-written-down-before-the-first-run",
   "heading": "What was written down before the first run",
   "level": 2,
   "span": [
    42410,
    54164
   ],
   "chunks": [
    [
     42410,
     54164
    ]
   ],
   "chars": 11754,
   "digest": "this section lists the pre-registration, the hardware used, the weights, the prompt, and the server settings. it also details the predictions made before the runs and the mistakes encountered during the bench.",
   "digest_skipped": null
  },
  {
   "id": "what-else-the-bench-read",
   "heading": "What else the bench read",
   "level": 2,
   "span": [
    54164,
    59302
   ],
   "chunks": [
    [
     54164,
     59302
    ]
   ],
   "chars": 5138,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-it-does-not-say",
   "heading": "What it does not say",
   "level": 2,
   "span": [
    59302,
    61354
   ],
   "chunks": [
    [
     59302,
     61354
    ]
   ],
   "chars": 2052,
   "digest": "the article specifies its limitations, noting it only used one card, one model family, and one operating system. it also clarifies that it does not compare other servers or production traffic.",
   "digest_skipped": null
  },
  {
   "id": "what-comes-next",
   "heading": "What comes next on this page",
   "level": 2,
   "span": [
    61354,
    62027
   ],
   "chunks": [
    [
     61354,
     62027
    ]
   ],
   "chars": 673,
   "digest": "the text outlines future additions to the page, including measurements of llama-server, different model weights, and whole-computer watts.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    62027,
    64528
   ],
   "chunks": [
    [
     62027,
     64528
    ]
   ],
   "chars": 2501,
   "digest": "the summary highlights that the two servers are close for one request, but vLLM's total speed reaches higher levels for many requests at once. it also notes that answers differ and memory usage depends on specific conditions.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work-and-see-it-live",
   "heading": "How to check our work — and see it live",
   "level": 2,
   "span": [
    64528,
    68865
   ],
   "chunks": [
    [
     64528,
     68865
    ]
   ],
   "chars": 4337,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    68865,
    70921
   ],
   "chunks": [
    [
     68865,
     70921
    ]
   ],
   "chars": 2056,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    70921,
    72191
   ],
   "chunks": [
    [
     70921,
     72191
    ]
   ],
   "chars": 1270,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    72191,
    76592
   ],
   "chunks": [
    [
     72191,
     76592
    ]
   ],
   "chars": 4401,
   "digest": null,
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-1454b854",
   "q": "how do ollama and vllm differ in memory management?",
   "a": "Ollama gives each request a slot that reserves a whole window of memory. vLLM batches continuously, handing out working memory in small pages as requests grow and reserving most of the card at start.",
   "cites": [
    "two-names-one-card"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-f5e7211e",
   "q": "how much memory does vllm reserve at rest?",
   "a": "vLLM 0.30.0 reserves 20,847 MiB on a cold boot and 22,505 MiB on a later boot that loaded a saved compile. This represents 84.8 per cent and 91.6 per cent of the card, respectively.",
   "cites": [
    "what-each-holds"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "flash-next-why-the-first-word-comes-sooner",
   "why": "shares ground with § Requests at once · § The desktop card: one engine a night"
  },
  {
   "slug": "ram-speeds-compared",
   "why": "shares ground with § The first word, and a shared prompt · § Why do some pages claim a 0.2-second first word?"
  },
  {
   "slug": "six-cards-the-same-questions",
   "why": "shares ground with § Requests at once · § Writing: how fast each one answers"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe six-way's 108 lines are our own, from [the long table](https://research.strata2signal.com/a-dinner-party-for-the-dead/), published with their scorer in the kit for Reading the answer. **Google** made Gemma 4 and its QAT weights and released them under the Apache 2.0 licence; **Hugging Face** hosts them and makes **transformers**, a library vLLM depends on. **Ollama**'s team and contributors built one server (MIT), on **llama.cpp** (MIT), begun by **Georgi Gerganov** and kept by the ggml authors, whose `llama-server` README states the defaults this page reads; the **vLLM** project (Apache-2.0), begun in UC Berkeley's **Sky Computing Lab** and hosted by the **PyTorch Foundation**, built the other, on **PyTorch** (BSD 3-clause). On this card both run over **CUDA** (NVIDIA, proprietary, the one closed piece in either stack). The card is an **EVGA** board on an **NVIDIA** chip. None of them owed us anything.\n\nA small human team asked for this page, set the rules and the predictions before the first run, chose what it would and would not claim, and signed off on every figure; a fleet of AI agents wrote the pre-registration, ran and checked each arm, and wrote this page under that team's rulings.",
 "kit": {
  "url": "https://research.strata2signal.com/ollama-and-vllm-on-one-rtx-3090/data/",
  "licence": "CC BY 4.0",
  "files": [
   "KIT-NOTE.md",
   "README.md",
   "kit/client/client.py",
   "kit/client/engine.py",
   "kit/client/engine_llamaserver.py",
   "kit/client/engine_ollama.py",
   "kit/client/framing.py",
   "kit/client/gates.py",
   "kit/client/nonces.py",
   "kit/client/parity.py",
   "kit/prompts/MANIFEST.json",
   "kit/prompts/frame/gemma4-canonical.json",
   "kit/prompts/long-prefill.txt",
   "kit/prompts/long-write.txt",
   "kit/prompts/prefix-questions.json",
   "kit/prompts/series-p512.txt",
   "kit/prompts/short-chat.txt",
   "kit/records/C-RESULTS.md",
   "kit/records/D-2026-09-30-RESULTS.md",
   "kit/records/P-RESULTS.md",
   "kit/rows/A/A.routes-cut.json",
   "kit/rows/C/C.power.csv",
   "kit/rows/C/RECEIPT",
   "kit/rows/C/parity.json",
   "kit/rows/D-2026-09-28/D.power.csv",
   "kit/rows/D-2026-09-28/RECEIPT",
   "kit/rows/D-2026-09-28/parity.json",
   "kit/rows/D-2026-09-30/D.power.csv",
   "kit/rows/D-2026-09-30/RECEIPT",
   "kit/rows/D-2026-09-30/parity.json",
   "kit/rows/P/P.power.csv",
   "kit/rows/P/RECEIPT",
   "kit/rows/P/parity.json",
   "kit/rows/Q/Q.power.csv",
   "kit/rows/Q/parity.json",
   "kit/rows/S/ovv-s-20260928T1130Z-ollama-scored-board.csv",
   "kit/rows/S/ovv-s-20260928T1130Z-ollama-scored.jsonl",
   "kit/rows/S/ovv-s-20260928T1130Z-vllm-scored-board.csv",
   "kit/rows/S/ovv-s-20260928T1130Z-vllm-scored.jsonl"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 13,
  "digests_kept": 8,
  "chips_written": 8,
  "chips_kept": 2,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 13
  },
  "new_noun_tokens_checked": [
   "0.30.0",
   "0.34.4",
   "000",
   "1",
   "1.2077",
   "1.284",
   "10",
   "108",
   "12B",
   "136.181",
   "142.327",
   "147.955",
   "16-bit",
   "17",
   "20",
   "212.999",
   "22",
   "299.14",
   "3.111",
   "3090",
   "4",
   "4-bit",
   "505",
   "506.829",
   "58.33",
   "63",
   "84.8",
   "847",
   "91.6",
   "97",
   "970.201",
   "C",
   "GGUF",
   "Gemma",
   "MiB",
   "Ollama",
   "Ollama's",
   "Q",
   "S",
   "W"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the board power while busy on the 3090 was 299.14 W",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 63,
    "text_sha": "51dc62ed4cc0",
    "figure": "212.999",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the line gives vLLM 0.30.0's 212.999 ms without its condition: arm C's reading of the first word at one request on the headline posture; the same page reads vLLM's first token from about 48 ms (a cached prompt, one request) to over 17 s (the 95th percentile, sixteen long prompts at once)",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its terminal figure). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the line gives vLLM 0.30.0's 212.999 ms without its condition: arm C's reading of the first word at one request on the headline posture; the same page reads vLLM's first token from about 48 ms (a cached prompt, one request) to over 17 s (the 95th percentile, sixteen long prompts at once). Not…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 203,
    "text_sha": "221b4ba6c179",
    "section": "two-names-one-card",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the digest makes the shared cache of prompt pages part of how vLLM batches continuously; the section gives continuous batching as small pages handed out as requests grow, and the prompt cache as how vLLM reuses a repeated prompt, a separate thing",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the digest makes the shared cache of prompt pages part of how vLLM batches continuously; the section gives continuous batching as small pages handed out as requests grow, and the prompt cache as how vLLM reuses a repeated prompt, a separate thing. Not a gate drop, and recorded here so the led…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 147,
    "text_sha": "4af145a07e8b",
    "section": "are-the-answers-the-same",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the digest's sentence on the 97 per cent can be read as the two disagreeing on 97 per cent of items; 97 per cent is the agreement the bet asked for and missed, and the digest drops what the page read instead: 63 of 108 items (58.33 per cent) on the strict scorer at one request, by the registered count, 54 of them items where neither reply opened with a letter",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the digest's sentence on the 97 per cent can be read as the two disagreeing on 97 per cent of items; 97 per cent is the agreement the bet asked for and missed, and the digest drops what the page read instead: 63 of 108 items (58.33 per cent) on the strict scorer at one request, by the registe…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 155,
    "text_sha": "b7ee239c596a",
    "section": "what-else-the-bench-read",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the digest lists the structured answers, the sandbox's network attempts and the runner's arguments as readings beside the routes; the section prints only the routes and says the other three join the page with arm G's rows in a dated update",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the digest lists the structured answers, the sandbox's network attempts and the runner's arguments as readings beside the routes; the section prints only the routes and says the other three join the page with arm G's rows in a dated update. Not a gate drop, and recorded here so the ledger add…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 144,
    "text_sha": "6d92e2ec8f46",
    "section": "sources",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the digest names llama.cpp's code among the sources; for llama.cpp the section cites its server README at build b11081, read beside the bundled program's own behaviour",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the digest names llama.cpp's code among the sources; for llama.cpp the section cites its server README at build b11081, read beside the bundled program's own behaviour. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 202,
    "text_sha": "6ff6accb4e8a",
    "chip_id": "c-85652149",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the answer gives 63 of 108 as items the two agreed on and drops the registered count's condition: an item on which neither reply opens with a letter counts as agreeing, and 54 of the 63 are such items, with 9 on which both opened with the same letter; it also drops that P-6 reads at one request",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its question). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the answer gives 63 of 108 as items the two agreed on and drops the registered count's condition: an item on which neither reply opens with a letter counts as agreeing, and 54 of the 63 are such items, with 9 on which both opened with the same letter; it also drops that P-6 reads at one reque…"
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 329,
    "text_sha": "c299f5bfc882",
    "chip_id": "c-b14156cc",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-07",
    "why": "the answer gives decode speed as how fast the answer is written, and total speed as every token written across all requests from sending to the last one finishing; the section's glossary gives decode speed as how fast one answer is written after the wait for the first token, and total speed as every token written across all requests over the time from sending to the last one finishing, so the answer loses that decode speed leaves that wait out and that total speed is a rate over that time, not a count",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-08: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-07 (matched on its question). The earlier record reads: removed from this pack by hand on 2026-10-07 (UTC), after the writer ran and after the gates had counted: the answer gives decode speed as how fast the answer is written, and total speed as every token written across all requests from sending to the last one finishing; the section's glossary gives decode speed as how fast one answer is written after the wait for the first token, and total speed a…"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 43,
    "text_sha": "1545193aa46c",
    "figure": "1.2077",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the line gives the 1.2077 ratio without its condition: vLLM total over the total of Ollama with sixteen slots, read at two requests at once on the 536-token prompt, 256 tokens out; it also drops that P-3 read UNDECIDED, the crossover coming at a level marked undecided before any run with a margin within run noise of the 1.2 bar, and that answers differ at that level",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the line gives the 1.2077 ratio without its condition: vLLM total over the total of Ollama with sixteen slots, read at two requests at once on the 536-token prompt, 256 tokens out; it also drops that P-3 read UNDECIDED, the crossover coming at a level marked undecided before any run with a margin within run noise of the 1.2 bar, and that answers differ at that level. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 206,
    "text_sha": "e8434fc5297f",
    "section": "one-request-at-a-time",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the digest gives a higher total speed for vLLM as the finding of arm S and drops its condition: one answer from send to last token on the proven pair, vLLM 0.27.1 and Ollama 0.32.13 with sixteen slots, a row the section says carries the load step of about 761 ms that Ollama 0.32.13 spent on each request and that reads about 2 ms on 0.34.4; the reading of record there is decode, within the 10 per cent band with Ollama faster, P-1 CONFIRMED",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the digest gives a higher total speed for vLLM as the finding of arm S and drops its condition: one answer from send to last token on the proven pair, vLLM 0.27.1 and Ollama 0.32.13 with sixteen slots, a row the section says carries the load step of about 761 ms that Ollama 0.32.13 spent on each request and that reads about 2 ms on 0.34.4; the reading of record there is decode, within the 10 per cent band with Ollama faster, P-1 CONFIRMED. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 302,
    "text_sha": "2e51da5d4d20",
    "chip_id": "c-0b8b8804",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the answer gives arm C totals without the prompt they were read on (the 536-token prompt, 256 tokens out) and prints the Ollama sixteen-slot totals at four and eight as steady readings, where the section says the sixteen-slot runs at four and at eight disagreed with each other past the P-3 band; it also drops the label every speed row carries, answers differ at this level; the reason a chip on this question left this pack on 2026-10-07",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the answer gives arm C totals without the prompt they were read on (the 536-token prompt, 256 tokens out) and prints the Ollama sixteen-slot totals at four and eight as steady readings, where the section says the sixteen-slot runs at four and at eight disagreed with each other past the P-3 band; it also drops the label every speed row carries, answers differ at this level; the reason a chip on this question left this pack on 2026-10-07. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 249,
    "text_sha": "b9c1a87d0084",
    "chip_id": "c-1e7e1100",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the answer gives the 1.284 to 3.111 ms load step without the condition TT read it under (one request, the window set at the server and none in the request; a request asking for a larger window reloads the model), and gives 10,970.201 ms without the condition the section prints it with: one request, read straight after the same runner had answered the same question sixteen at a time with no restart between, a condition and not a cause, beside 432.838 ms on the one-slot default",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the answer gives the 1.284 to 3.111 ms load step without the condition TT read it under (one request, the window set at the server and none in the request; a request asking for a larger window reloads the model), and gives 10,970.201 ms without the condition the section prints it with: one request, read straight after the same runner had answered the same question sixteen at a time with no restart between, a condition and not a cause, beside 432.838 ms on the one-slot default. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 248,
    "text_sha": "c972c6809104",
    "chip_id": "c-3c22c396",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the answer sets a 4-bit GGUF against a 16-bit output layer and drops the other half of the comparison the section makes: the linear layers are 4-bit in blocks of 32 on both, and the output layer is 6-bit in the GGUF against 16-bit in the vLLM checkpoint, which is where nearly all of the 17 per cent lies",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the answer sets a 4-bit GGUF against a 16-bit output layer and drops the other half of the comparison the section makes: the linear layers are 4-bit in blocks of 32 on both, and the output layer is 6-bit in the GGUF against 16-bit in the vLLM checkpoint, which is where nearly all of the 17 per cent lies. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 312,
    "text_sha": "df07770cff5d",
    "chip_id": "c-157535e3",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-08",
    "why": "the answer gives the vLLM routes as a finding and drops the label the section puts on them: they are routes the vLLM security guide at the tag names as served without the key, each called on the loopback address only, and the table claims nothing about any route it does not list; it also counts five Ollama routes as open without a key, where the section says Ollama has no key system and lists five routes, none of which adds, changes or removes a model; the reason a chip on this question left this pack on 2026-10-07",
    "detail": "removed from this pack by hand on 2026-10-08 (UTC), after the writer ran and after the gates had counted: the answer gives the vLLM routes as a finding and drops the label the section puts on them: they are routes the vLLM security guide at the tag names as served without the key, each called on the loopback address only, and the table claims nothing about any route it does not list; it also counts five Ollama routes as open without a key, where the section says Ollama has no key system and lists five routes, none of which adds, changes or removes a model; the reason a chip on this question left this pack on 2026-10-07. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-10-08T02:46:41Z"
}