{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "the-llama-weights-that-got-out",
 "title": "The weights that got out: how LLaMA started local AI",
 "dek": "Meta sent LLaMA to approved researchers in February 2023; within a week the files were on BitTorrent, within three llama.cpp, Alpaca and Vicuna had been built on them. The leak, the Senate letter, the Llama 2 licence the Open Source Initiative called not open source, every Llama since, dated, and why the phrase became 'open weights' — now that Meta's newest open weights carry another name.",
 "published": "2026-11-04",
 "series": [
  "notes"
 ],
 "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/the-llama-weights-that-got-out/",
  "md_url": "https://research.strata2signal.com/the-llama-weights-that-got-out/index.md",
  "md_sha": "e7256c103f0127d1802d2d21eb8662e63c2dca2f7f47001026d4f2ce78c829ba",
  "html_sha": "355a9a53c573f0dbe14331b30163aa2273f340f61b65aa6daa2df8ff38152583",
  "anchors_sha": "21df53b22c6dd3b71485c97d82780fe9053701d896cfb266370d70cff6980749"
 },
 "short": {
  "paragraph": "On 2023-02-24 Meta announced LLaMA, four language models of 7 to 65 billion parameters, for researchers by application under a non-commercial licence. By 2023-03-02 the files were circulating by torrent. On 2023-03-10 Georgi Gerganov published llama.cpp, which ran them on ordinary computers; on 2023-03-13 Stanford's Alpaca fine-tuned the 7B for under $600; on 2023-03-30 Vicuna followed. On 2023-07-18 Meta released Llama 2 \"free for research and commercial use\", under a licence the Open Source Initiative said two days later is not open source. Llama 3, 3.1, 3.2 and 3.3 (2024) and Llama 4 (2025-04-05) followed under the same shape of licence; Meta's own word for Llama 4 was \"open-weight\", and on 2026-11-04 Llama 4 is still the newest Llama in Meta's own records (its Hugging Face account and its table of launch dates); the open weights Meta has shipped since include a new family under another name, Muse: a 30B model under Apache 2.0 on 2026-08-10 and an 816B model under a licence of Meta's own on 2026-10-15. This page measures nothing; the Llama-family readings this workshop has published are listed at the end.",
  "from_dek": false,
  "counts": {
   "words": 4411,
   "minutes": 20,
   "tables": 1,
   "kit": false
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "why-this-page",
   "heading": "Why this page",
   "level": 2,
   "span": [
    3086,
    3636
   ],
   "chunks": [
    [
     3086,
     3636
    ]
   ],
   "chars": 550,
   "digest": "the models at this workshop's doors run through Ollama over llama.cpp, which began as a program for running the LLaMA files on a laptop. this page covers the month before Ollama and the licences after.",
   "digest_skipped": null
  },
  {
   "id": "february-2023",
   "heading": "February 2023: a model for researchers, by application",
   "level": 2,
   "span": [
    3636,
    5202
   ],
   "chunks": [
    [
     3636,
     5202
    ]
   ],
   "chars": 1566,
   "digest": "on 2023-02-24 Meta announced LLaMA, a foundational, 65-billion-parameter large language model, in four sizes. the model was released under a noncommercial license focused on research use cases for academics, people in government and civil society, and industry research laboratories.",
   "digest_skipped": null
  },
  {
   "id": "the-first-week",
   "heading": "The first week: the files got out",
   "level": 2,
   "span": [
    5202,
    6324
   ],
   "chunks": [
    [
     5202,
     6324
    ]
   ],
   "chars": 1122,
   "digest": "six days after the announcement, a pull request to Meta's repository proposed using a torrent to distribute the model. a letter from two United States senators recorded that the full model appeared on BitTorrent, making it available to anyone, anywhere in the world, without monitoring or oversight.",
   "digest_skipped": null
  },
  {
   "id": "three-weeks",
   "heading": "Three weeks: llama.cpp, Alpaca, Vicuna",
   "level": 2,
   "span": [
    6324,
    9090
   ],
   "chunks": [
    [
     6324,
     9090
    ]
   ],
   "chars": 2766,
   "digest": "llama.cpp aimed to run the model using 4-bit quantization on a MacBook. Stanford's Alpaca was a model fine-tuned from the LLaMA 7B model on 52K instruction-following demonstrations. Vicuna-13B was trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT.",
   "digest_skipped": null
  },
  {
   "id": "july-2023",
   "heading": "July 2023: Llama 2, and a licence that is not open source",
   "level": 2,
   "span": [
    9090,
    10601
   ],
   "chunks": [
    [
     9090,
     10601
    ]
   ],
   "chars": 1511,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-llama-line",
   "heading": "2024 to 2025: the Llama line, dated",
   "level": 2,
   "span": [
    10601,
    13655
   ],
   "chunks": [
    [
     10601,
     13655
    ]
   ],
   "chars": 3054,
   "digest": "the Llama line includes Llama 3, Llama 3.1, Llama 3.2, Llama 3.3, and Llama 4. Meta's Llama 4 includes Scout and Maverick models. since 2026-08-10, Meta has also shipped open weights under the Muse name, including Muse Glimmer and Muse Spark Open.",
   "digest_skipped": null
  },
  {
   "id": "why-open-weights",
   "heading": "Why the phrase is \"open weights\"",
   "level": 2,
   "span": [
    13655,
    15124
   ],
   "chunks": [
    [
     13655,
     15124
    ]
   ],
   "chars": 1469,
   "digest": "Open Weights refer to the final weights and biases of a trained neural network, excluding the training code and dataset. Meta's terminology moved from a noncommercial license to open-weight by 2025.",
   "digest_skipped": null
  },
  {
   "id": "what-this-shelf-has-measured",
   "heading": "What this shelf has measured on Llama",
   "level": 2,
   "span": [
    15124,
    17304
   ],
   "chunks": [
    [
     15124,
     17304
    ]
   ],
   "chars": 2180,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    17304,
    18497
   ],
   "chunks": [
    [
     17304,
     18497
    ]
   ],
   "chars": 1193,
   "digest": "the first LLaMA was never released to the public. local AI's tools like llama.cpp, Alpaca, and Vicuna were built for those files in weeks. Llama 2 made it legal under a licence of Meta's own. \"open weights\" names what is given.",
   "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": [
    18497,
    19503
   ],
   "chunks": [
    [
     18497,
     19503
    ]
   ],
   "chars": 1006,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    19503,
    20534
   ],
   "chunks": [
    [
     19503,
     20534
    ]
   ],
   "chars": 1031,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    20534,
    21820
   ],
   "chunks": [
    [
     20534,
     21820
    ]
   ],
   "chars": 1286,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    21820,
    33881
   ],
   "chunks": [
    [
     21820,
     33881
    ]
   ],
   "chars": 12061,
   "digest": "this section provides links and details for every announcement, licence, and model card mentioned in the article, including dates and specific details from Meta, arXiv, GitHub, and other sources.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-f34b22b9",
   "q": "how did meta describe the initial llama release?",
   "a": "Meta announced LLaMA as a foundational, 65-billion-parameter large language model in four sizes, intended for research use cases by those without large amounts of infrastructure.",
   "cites": [
    "february-2023"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-7cd68f1c",
   "q": "how did the model files become widely available?",
   "a": "Within days of the announcement, the full model appeared on BitTorrent, making it available to anyone without monitoring or oversight.",
   "cites": [
    "the-first-week"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-9ac4fb8e",
   "q": "what were the goals of llama.cpp?",
   "a": "The main goal was to run the model using 4-bit quantization on a MacBook to store weights in fewer bits.",
   "cites": [
    "three-weeks"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-f5d29ad8",
   "q": "what distinguishes open weights from open source?",
   "a": "Open weights refer to the final weights and biases of a trained neural network, whereas open source requires the training code, parameters, and detailed information about the training data.",
   "cites": [
    "why-open-weights"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a359bd0a",
   "q": "how did llama3.3:70b perform in seat trials?",
   "a": "The model threw out all 27 claims that were not in their source and kept 13 of the 16 that were, resulting in a PASS with a warm median of 1,386 ms.",
   "cites": [
    "what-this-shelf-has-measured"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-aabce1d0",
   "q": "what key points summarize the llama history?",
   "a": "The first LLaMA was never released to the public, local AI tools were built for those files in weeks, and Llama 2 introduced a specific commercial licence.",
   "cites": [
    "what-to-take-with-you"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "a-short-history-of-ollama",
   "why": "shares ground with § Why this page · § The engine underneath: llama.cpp"
  },
  {
   "slug": "a-short-history-of-qwen",
   "why": "shares ground with § Why the phrase is \"open weights\" · § What to take with you"
  },
  {
   "slug": "how-we-work",
   "why": "shares ground with § Why this page · § The plain-english version"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to **Meta** and the authors of the LLaMA and Llama 2 papers, whose announcements, model cards and licence texts are dated and still up, which is why this page could be written from primary sources rather than from memory; to **Georgi Gerganov** and **the ggml authors** for **llama.cpp** (MIT), the engine at the bottom of most of this shelf's language-model numbers; to the **Stanford Center for Research on Foundation Models** for the Alpaca page and to **LMSYS** for the Vicuna page, both still carrying their dates and their terms; to **Simon Willison**, whose post of 2023-03-11 is the one first-hand account of that week we could read in full; to the **Open Source Initiative** for the Llama 2 post, the Open Source AI Definition, its FAQ and its open-weights page; to **Senators Richard Blumenthal and Josh Hawley**, whose letter of 2023-06-06 is public on the Senate's site; and to **GitHub** and **Hugging Face**, whose records gave the repository, commit and model dates. None of them owed us anything. A small human team asked for this page, chose what it would and would not claim, and signed off on it; a fleet of AI agents fetched and read every source listed below and drafted it under that team's rulings.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 9,
  "digests_kept": 8,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 1,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 5
  },
  "new_noun_tokens_checked": [
   "1",
   "13",
   "16",
   "2",
   "2023-02-24",
   "2025",
   "2026-08-10",
   "27",
   "3",
   "3.1",
   "3.2",
   "3.3",
   "386",
   "4",
   "4-bit",
   "405B",
   "52K",
   "65-billion-parameter",
   "70b",
   "7B",
   "8b",
   "AI",
   "AI's",
   "Alpaca",
   "Behemoth",
   "BitTorrent",
   "GitHub",
   "Glimmer",
   "LLaMA",
   "Llama",
   "MacBook",
   "Maverick",
   "Meta",
   "Meta's",
   "Muse",
   "Ollama",
   "Open",
   "PASS",
   "Scout",
   "ShareGPT",
   "Spark",
   "States",
   "United",
   "Vicuna",
   "Vicuna-13B",
   "Weights",
   "llama3.1",
   "llama3.3"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "the Llama 3.1 release included a model with 405B",
    "detail": "405b is in the cited section only as code or in a table cell"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 53,
    "text_sha": "9db9b6041399",
    "figure": "4",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "it cuts the README's sentence at its figure: 'to run the model using 4-bit' drops 'quantization on a MacBook', so the line says neither what was 4-bit nor what it was to run on, and the laptop is the point the section makes",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: it cuts the README's sentence at its figure: 'to run the model using 4-bit' drops 'quantization on a MacBook', so the line says neither what was 4-bit nor what it was to run on, and the laptop is the point the section makes. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 83,
    "text_sha": "a193eb56a355",
    "figure": "65",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "it is not a sentence ('one with 65-billion-parameter'), and it reads the announcement's title adjective as one of the sizes; the section gives the four sizes as 7B, 13B, 33B and 65B parameters",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: it is not a sentence ('one with 65-billion-parameter'), and it reads the announcement's title adjective as one of the sizes; the section gives the four sizes as 7B, 13B, 33B and 65B parameters. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 210,
    "text_sha": "4d78ccf069fe",
    "section": "what-this-shelf-has-measured",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "it says this shelf measured 'several Llama models, including' two, where the section lists exactly two Ollama tags, the 70B and the 8B; and 'processor speed' names the machine as what was measured, where the page gives the 70B tag's own speed on processors alone, 2.7 tokens a second",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: it says this shelf measured 'several Llama models, including' two, where the section lists exactly two Ollama tags, the 70B and the 8B; and 'processor speed' names the machine as what was measured, where the page gives the 70B tag's own speed on processors alone, 2.7 tokens a second. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 144,
    "text_sha": "c56defc15d72",
    "chip_id": "c-7aca6874",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "it turns the section's 'the month before it' into 'the month preceding the release of Ollama', a dated claim the page does not make: the page's month is February to March 2023, and Ollama's first release, which the page dates from the Ollama history, was 2023-07-08",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: it turns the section's 'the month before it' into 'the month preceding the release of Ollama', a dated claim the page does not make: the page's month is February to March 2023, and Ollama's first release, which the page dates from the Ollama history, was 2023-07-08. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 215,
    "text_sha": "7c0704cb6db9",
    "chip_id": "c-5ed6c3a6",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "it counts Behemoth among 'the first open-weight natively multimodal models'; the section's table gives Scout and Maverick as the release and Behemoth as 'shown still training'",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: it counts Behemoth among 'the first open-weight natively multimodal models'; the section's table gives Scout and Maverick as the release and Behemoth as 'shown still training'. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-11-04T12:17:59Z"
}