The weights that got out

If you're new here: strata→signal is a small workshop (plus a friendly dog with a white patch) that builds on its own machines and writes up what it measures. Every date below was read on 2026-10-14 and read again on 2026-10-29 and 2026-11-04 (UTC) from the source credited beside it — Meta's announcements, model cards and licence texts; the projects' repositories and GitHub's own record of them; the Stanford and LMSYS posts; the Open Source Initiative's pages; a letter from two United States senators; and one blog post written the week it happened. Where we found no source, the claim is not here. Nothing on this page links to, or says how to get, the files that leaked.

Two words this page leans on. A model's weights are the numbers training produced: billions of them, in files, and the whole of what a model is once the code to run them exists. Open weights are weights anyone may download, under terms that vary by licence; the last section is about the phrase.

Why this page

The models at this workshop's doors — the Mistral model that reads what strangers type at RuleSage, the Gemma model that narrates RealKeep — run through Ollama over llama.cpp, and llama.cpp began as a program for one thing: running the LLaMA files on a laptop. A short history of Ollama tells the server's story; this page is the month before it, and the licences after.

February 2023: a model for researchers, by application

On 2023-02-24 Meta announced LLaMA, "a foundational, 65-billion-parameter large language model", in four sizes: 7B, 13B, 33B and 65B parameters, the two largest trained on 1.4 trillion tokens. The reason it gave was computing power: smaller models let "others in the research community who don't have access to large amounts of infrastructure" study them. The terms were in the same post: "we are releasing our model under a noncommercial license focused on research use cases. Access to the model will be granted on a case-by-case basis" to academics, to people in government and civil society, and to industry research laboratories.

The paper went to arXiv on 2023-02-27 with fourteen authors, two of whom, Guillaume Lample and Timothée Lacroix, would found Mistral AI weeks later (A short history of Mistral). Its abstract claimed "LLaMA-13B outperforms GPT-3 (175B) on most benchmarks", and said "We release all our models to the research community".

The repository, created on GitHub on 2023-02-14, took its first commit from Lacroix on 2023-02-24. Its README sent applicants to a form: "Once your request is approved, you will receive links to download the tokenizer and model files". The code was under the GNU GPL version 3; the weights were not. The model card said "Non-commercial bespoke license", named researchers as the primary intended users, and dated the training "between December. 2022 and Feb. 2023".

The first week: the files got out

On 2023-03-02 at 23:05 UTC, six days after the announcement, a pull request to Meta's own repository proposed to "Save bandwidth by using a torrent to distribute more efficiently". It was closed unmerged, and it is not linked here. A letter from two United States senators that June put the week in one sentence: "within days of the announcement, the full model appeared on BitTorrent, making it available to anyone, anywhere in the world, without monitoring or oversight."

Simon Willison, a programmer who had been writing about these models, wrote on 2023-03-11: "So they're in the wild now. You may not be legally able to build a commercial product on them, but the genie is out of the bottle." The senators' letter records both of Meta's positions at once: "While Meta has described the release as a leak, its chief AI scientist has stated that open models are key to its commercial success." Richard Blumenthal and Josh Hawley asked Meta nine questions, among them how many researchers it had approved; Meta's answer, if it gave one, is in no source we read.

Three weeks: llama.cpp, Alpaca, Vicuna

2023-03-10. Georgi Gerganov's first commit to llama.cpp, "Initial release", is stamped 18:40 UTC; GitHub created the repository that day; its MIT licence file followed on 2023-03-12. Its README, added at 19:47 UTC on 2023-03-10, gave the aim: "The main goal is to run the model using 4-bit quantization on a MacBook" — storing the weights in fewer bits so they fit (The compressed photograph explains the trick). Simon Willison, writing the next day, ran the 7B on his own laptop, then the 13B: "I thought it would be a few more years before I could run a GPT-3 class model on hardware that I owned. I was wrong: that future is here already." The rest of llama.cpp's story — GGUF in August 2023, the engine under Ollama from that July — is on the Ollama page.

2023-03-13. Stanford's Center for Research on Foundation Models announced Alpaca, "a model fine-tuned from the LLaMA 7B model on 52K instruction-following demonstrations" — a fine-tune is a model trained a little further on new examples — the demonstrations generated with OpenAI's text-davinci-003 for "less than $500", the fine-tune done for "less than $100". It was "intended only for academic research", first because of LLaMA's non-commercial licence, second because OpenAI's terms forbid "developing models that compete with OpenAI", third because its safety measures were not yet adequate for general use. The data and code were released; the weights were held back pending Meta's guidance. The page now says "The public demo is now disabled".

2023-03-30. A team from UC Berkeley, CMU, Stanford, UC San Diego and MBZUAI announced Vicuna-13B, "trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT" for "around $300", claiming "more than 90%* quality of OpenAI ChatGPT and Google Bard" with GPT-4 as the judge; the asterisk is the post's own: "According to a fun and non-scientific evaluation with GPT-4. Further rigorous evaluation is needed." Its demo was "a research preview intended for non-commercial use only", subject to LLaMA's licence; the code was Apache 2.0. For the two months before vLLM's own announcement on 2023-06-20, by vLLM's account, the server behind that demo was vLLM (A short history of vLLM).

Alpaca and Vicuna each inherited LLaMA's terms, and each said so: nothing built on the files could be sold. llama.cpp was different in kind: its own code took the MIT licence, which allows selling, and it shipped no weights; the files it ran carried Meta's terms.

July 2023: Llama 2, and a licence that is not open source

On 2023-07-18 Meta announced Llama 2: "Llama 2 is free for research and commercial use", with Microsoft "as the preferred partner", and said it had been "blown away by the huge demand for Llama 1 from researchers — with more than 100,000 requests for access". The paper, 68 authors, went to arXiv the same day: models "ranging in scale from 7 billion to 70 billion parameters", and a dialogue version, Llama 2-Chat. Meta's post said "We're now ready to open source the next version of Llama 2". Ollama had shipped its first release ten days earlier.

The "LLAMA 2 COMMUNITY LICENSE AGREEMENT", release date "July 18, 2023", has two clauses that mattered: a licensee whose products had, on the Llama 2 release date, "greater than 700 million monthly active users in the preceding calendar month" "must request a license from Meta", and "You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof)". An Acceptable Use Policy is incorporated by reference. On 2023-07-20 Stefano Maffulli of the Open Source Initiative wrote that the licence "is not Open Source": "it puts restrictions on commercial use for some users (paragraph 2)" and "restricts the use of the model and software for certain purposes", failing the Open Source Definition's points 5 and 6, no discrimination against persons, groups or fields of endeavour.

2024 to 2025: the Llama line, dated

Mark Zuckerberg's letter of 2024-07-23, "Open Source AI Is the Path Forward", gave the business reason: "selling access to AI models isn't our business model. That means openly releasing Llama doesn't undercut our revenue". By 2024-12-19 Meta counted "more than 650 million downloads of Llama and its derivatives".

After April 2025 the Llama record stops. The newest repositories under Meta's meta-llama account on Hugging Face, read on 2026-10-14, 2026-10-29 and 2026-11-04, are the Llama 4 weights (2025-04-01 to 2025-04-04) and three safety models from the same month, and its own table of launch dates ends at Llama 4. The open weights Meta has shipped since include a new family under another name, Muse, which sits under a second account, meta-models. On 2026-08-10 it announced Muse Glimmer, "a 30-billion-parameter model optimized for always-on local agent workflows", and said it was "open sourcing the model weights under a permissive Apache 2.0 license". We have run Muse Glimmer on our own cards: The August arrivals and the chair trials give its readings, and say where a reading could not be scored. On 2026-10-15 it released Muse Spark Open, "816B total parameters (42B active)", "the most capable US open-weight model to date" in its own words, "under the Muse Model License with limited commercial terms"; we have not read that licence, and neither post uses the word Llama. Its developer site on 2026-11-04 leads with the Muse family and names Llama only in a menu and its footer. Whether another Llama is coming, no source we read says.

Why the phrase is "open weights"

The Open Source Initiative's own page defines the term: "Open Weights refer to the final weights and biases of a trained neural network", and says what they leave out: the training code and the training dataset. Its Open Source AI Definition, version 1.0, announced on 2024-10-28, asks for the code and the parameters and, in place of the dataset itself, "sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system", all under terms that let anyone use, study, modify and share the system; its FAQ lists "Llama2 (Meta)" among systems it analysed that do not comply, because they "lack required components and/or their legal agreements are incompatible with the Open Source principles".

Meta's own words moved over the same three years: "a noncommercial license focused on research use cases" (2023-02-24); "open source" (2023-07-18); "openly available" (2024-04-18); "open source AI model" (2024-07-23); "open-weight" (2025-04-05); and, for the Muse models, "open sourcing the model weights" (2026-08-10) and "open-weight" again (2026-10-15). The phrase that names exactly what is handed over — the weights, under whatever licence comes with them — is the one Meta itself reached by 2025 and still used in 2026, and the one the Open Source Initiative defines. This page does not say who said it first; no source we read does.

What this shelf has measured on Llama

None of this workshop's products runs a Llama model. The readings it has published on the family, each on its own page with its conditions:

• llama3.3:70b (an Ollama tag, the name that pulls one build of a model), 70.6B at 4-bit, sat the seat trials (published 2026-08-11), the exam a model passes before it may judge claims inside our products: it threw out all 27 claims that were not in their source and kept 13 of the 16 that were (kills /27: 27, keeps /16: 13), PASS, warm median 1,386 ms; the same row is carried on The August arrivals among what that page calls "the July field".
• The same tag on processors alone, no graphics card in the path, 2026-08-25: 2.7 tokens a second with 41 GiB resident, in the server column of Two Hours at 12 tok/s — On Battery, median of three runs, and 37.6 s to the first token on a 505-token prompt, nothing cached.
• The same tag as one of eight arms in the llms.txt reading, measured 2026-08-16: 1 of 24 questions answered right with no context, 10 of 20 with the map, 7 of 20 with the site's own pages (its C-NONE, C-MAP and C-HTML columns).
• llama3.1:8b in the voice trials (published 2026-08-11), in its first era, the 8 GB-card rig, a round that page calls "a practitioner's log", judged by eye: loaded ~5.0 GB, warm turn ~5 s, verdict "deflects into meta — despite persona-benchmark fame".

The licence ledger carries no Llama row: this shelf has never read a Llama licence off a tag it ran, and says so. Before this page, the one Llama licence document it had read was Meta's "Llama 3 Acceptable Use Policy", for its title only, on 2026-10-02, for Uncensored models; the licence texts quoted above were read for this page from Meta's repositories, not off a tag we ran.

What to take with you

• The first LLaMA was never released to the public. Approved researchers, a bespoke non-commercial licence, 2023-02-24; circulating anyway by 2023-03-02.
• Local AI's tools were built for those files, in weeks. llama.cpp 2023-03-10, Alpaca 2023-03-13, Vicuna 2023-03-30. llama.cpp's own code was MIT-licensed; Alpaca and Vicuna, made from the weights, carried LLaMA's terms, which forbade selling anything built on them.
• Llama 2 made it legal, under a licence of Meta's own. "Free for research and commercial use" from 2023-07-18, with a 700-million-user clause and a use policy; not open source by the Open Source Initiative's reading of 2023-07-20, nor by its 2024 definition.
• "Open weights" names what is given. The weights, under their licence, without the data or the training code. Meta's own word for Llama 4 was "open-weight", and on 2026-11-04 Llama 4 is the newest Llama in Meta's own records. Since 2026-08-10 Meta has shipped open weights under the Muse name, not as Llama: Muse Glimmer, 30B, under Apache 2.0 that day, and Muse Spark Open on 2026-10-15, under the Muse Model License, which we have not read.

How to check our work — and see it live

• Read the dates where they live. Every announcement, licence and model card on this page is linked under Sources, and each carries its own date; the two Meta sources that disagree on Llama 3.3 are one click apart.
• Ask GitHub yourself. With GitHub's command-line tool, gh api repos/ggml-org/llama.cpp/commits/26c0846 prints the "Initial release" commit at 2023-03-10T18:40:58Z, and gh api repos/meta-llama/llama prints the repository's creation, 2023-02-14T09:29:12Z.
• Read the licences side by side. The Llama 2, 3 and 4 licences are linked under Sources; search each for "700 million", and the Llama 3 and 4 texts for "Built with", a clause the Llama 2 text does not have.
• See the engine at work. RuleSage is free and needs no account; the Mistral model that reads questions there runs through Ollama over llama.cpp, the program this page watched begin.

The rest of the seminar

• A short history of Ollama: the server most of this shelf's numbers ran through.
• A short history of vLLM: the server behind the Vicuna demo.
• A short history of Mistral: two LLaMA authors' company, and its open-weight releases.
• A short history of Qwen: the same shape for Alibaba's models, whose licences split by size.
• The compressed photograph: what 4-bit quantization does to a model.
• Uncensored models: fine-tunes that change the weights and not the licence.
• The licence ledger: every licence this shelf has read first-hand, dated.

Who ran this, and thanks

Thanks 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.

Sources

Every external link below was fetched and read on 2026-10-14 between 01:16 and 01:25 (UTC), and re-read for the page's check and its fix between 06:18 and 06:38 (UTC); every link was tried again on 2026-10-29 between 16:21 and 16:28 (UTC) and on 2026-11-04 between 11:39 and 11:47 (UTC), when one link was added (a second copy of the senators' letter), and the counts that moved carry the later date. Repository, commit and pull-request dates are GitHub's record, read through its API; the pull request that offered the leaked files was read for its title, date and state only, and is not linked.

• Meta — Introducing LLaMA: A foundational, 65-billion-parameter large language model, 2023-02-24 (the four sizes; "We trained LLaMA 65B and LLaMA 33B on 1.4 trillion tokens"; the "noncommercial license" and case-by-case sentences; "Smaller, more performant models such as LLaMA enable others in the research community who don't have access to large amounts of infrastructure to study these models")
• arXiv — LLaMA: Open and Efficient Foundation Language Models, v1 submitted 2023-02-27 17:11 UTC; fourteen authors, Lacroix and Lample among them; the two abstract sentences quoted
• GitHub — meta-llama/llama, formerly facebookresearch/llama (created 2023-02-14T09:29Z; "Initial commit" 2023-02-24T13:20Z by Timothée Lacroix; 59,628 stars on 2026-11-04; its README now carries a "Note of deprecation": "As part of the Llama 3.1 release, we've consolidated GitHub repos"); at the initial commit, its README (the request form; "Once your request is approved…") and LICENSE (GNU GPL version 3); at the commit of 2023-03-02, its MODEL_CARD.md ("Non-commercial bespoke license"; "researchers in natural language processing, machine learning and artificial intelligence"; "trained between December. 2022 and Feb. 2023"); the pull request of 2023-03-02T23:05Z, closed and not merged
• GitHub — ggml-org/llama.cpp (created 2023-03-10T18:58Z; the "Initial release" commit by Georgi Gerganov, 2023-03-10T18:40Z; its README, created 2023-03-10T19:47Z ("The main goal is to run the model using 4-bit quantization on a MacBook"); its LICENSE (MIT), added 2023-03-12T06:36Z; "LLM inference in C/C++"; 133,338 stars on 2026-11-04)
• Simon Willison — Large language models are having their Stable Diffusion moment, 2023-03-11 (the pull request; "in the wild"; "the genie is out of the bottle"; the "initial release yesterday"; the README's "main goal" sentence, quoted; the 7B then the 13B on his laptop; "I was wrong: that future is here already")
• United States Senate — the letter of Senators Richard Blumenthal and Josh Hawley to Mark Zuckerberg, 2023-06-06 (a five-page PDF; "within days of the announcement, the full model appeared on BitTorrent…"; "While Meta has described the release as a leak…"; nine questions, question 1a on how many researchers were approved; signed as Chair and Ranking Member of the Subcommittee on Privacy, Technology, & the Law); on 2026-11-04 the PDF turned our automated reader away, so both quoted sentences and question 1a were checked again, word for word, and the nine questions counted again, against the letter's full text in Senator Hawley's release of the same day
• Stanford CRFM — Alpaca: A Strong, Replicable Instruction-Following Model, 2023-03-13 (the sentences quoted, "less than $500 using the OpenAI API" and "less than $100 on most cloud compute providers" among them; the three reasons; the weights held back; "The public demo is now disabled"); its repository tatsu-lab/stanford_alpaca (created 2023-03-10T23:33Z; Apache-2.0)
• LMSYS — Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality, 2023-03-30 (the sentences quoted; GPT-4 as judge; the asterisk's footnote; "non-commercial use only"; the code under Apache 2.0); its repository lm-sys/FastChat (created 2023-03-19T00:18Z; Apache-2.0)
• Meta — Meta and Microsoft Introduce the Next Generation of Llama, 2023-07-18 ("free for research and commercial use"; "preferred partner"; "more than 100,000 requests for access"; "We're now ready to open source the next version of Llama 2")
• arXiv — Llama 2: Open Foundation and Fine-Tuned Chat Models, v1 submitted 2023-07-18 14:31 UTC, 68 authors in arXiv's listing ("ranging in scale from 7 billion to 70 billion parameters"; "Llama 2-Chat")
• GitHub — the Llama 2 Community License ("Llama 2 Version Release Date: July 18, 2023"; the 700 million clause; the "improve any other large language model" clause; the Acceptable Use Policy by reference) and the Llama 2 model card ("A custom commercial license"; "intended for commercial and research use in English")
• Open Source Initiative — Stefano Maffulli, Meta's LLaMa license is not Open Source, 2023-07-20
• Meta — Introducing Meta Llama 3, 2024-04-18 ("8B and 70B parameters"; "over 15T tokens"; "the most capable openly available LLM to date"); Introducing Llama 3.1, 2024-07-23 (8B, 70B, 405B; 128K; "the first frontier-level open source AI model"; the licence change on outputs; "more than 300 million total downloads"); Mark Zuckerberg, Open Source AI Is the Path Forward, 2024-07-23 ("selling access to AI models isn't our business model…"); Llama 3.2, 2024-09-25 ("11B and 90B vision LLMs"; "1B and 3B"; "openness drives innovation…"); The future of AI: Built with Llama, 2024-12-19 ("more than 650 million downloads of Llama and its derivatives"); The Llama 4 herd, 2025-04-05 (Scout, Maverick and Behemoth as quoted; "first open-weight natively multimodal models")
• GitHub — meta-llama/llama-models (created 2024-06-27): its README's launch-date table, Llama 2 (7/18/2023) through Llama 4 (4/5/2025), with Llama 3.3 at "12/04/2024"; the Llama 3 licence ("META LLAMA 3 COMMUNITY LICENSE AGREEMENT"; "April 18, 2024"; the 700 million clause; "Built with Meta Llama 3"; the name clause); the Llama 4 licence ("LLAMA 4 COMMUNITY LICENSE AGREEMENT"; "Llama 4 Version Effective Date: April 5, 2025"; the 700 million clause; "Built with Llama"; the name clause)
• Hugging Face — meta-llama/Llama-3.3-70B-Instruct (model card: "December 6, 2024"; "Llama 3.3 Community License Agreement"; 70B; 128k), and the Hub's API listing of the meta-llama account's repositories by creation date, newest first, read 2026-10-14, 2026-10-29 and 2026-11-04, unchanged (the thirteen newest: ten Llama 4 repositories 2025-04-01 to 2025-04-04, Llama Guard 4 2025-04-23, two Prompt Guard 2 sizes 2025-04-28); the same listing for the meta-models account, read 2026-10-29 and again 2026-11-04, nothing newer (Muse-Glimmer-30B created 2026-08-09, licence tag apache-2.0; Muse-Spark-Open-816B-A42B created 2026-09-15, licence tag other, its card naming the "Muse Model License" and "Model Release Date: October 2026", access gated)
• Meta — the developer site, read 2026-10-14, 2026-10-29 and 2026-11-04 (llama.com redirects to it in two hops on all three days; on 2026-10-29 and again on 2026-11-04 Muse Wonder, Muse Wonder Full, Muse Spark 1.3, 1.2 and 1.1, Muse Spark Open, Muse Glimmer, Muse Image, Muse Voice Transcribe, Muse Code and SAM 3.1 named; Llama as "Llama ↗" in a Community menu and in the footer as "Llama Docs" and "Llama GitHub"); Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device, 2026-08-10 ("a 30-billion-parameter model optimized for always-on local agent workflows"; "open sourcing the model weights under a permissive Apache 2.0 license"; "you can download the weights on Hugging Face"); Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date, 2026-10-15 ("Today, we're releasing the weights for Muse Spark Open, the most capable US open-weight model to date"; "816B total parameters (42B active)"; "under the Muse Model License with limited commercial terms"; the word Llama in neither post)
• Open Source Initiative — The Open Source AI Definition, version 1.0 (the four freedoms; data information, code and parameters); its announcement, 2024-10-28, at All Things Open 2024; its FAQ ("Llama2 (Meta), Grok (X/Twitter), Phi-2 (Microsoft), and Mixtral (Mistral)"; "lack required components and/or their legal agreements are incompatible with the Open Source principles"); Open Weights: not quite what you've been told, undated ("Open Weights refer to the final weights and biases of a trained neural network")
• This workshop's own released pages, linked above and read in their markdown copies on 2026-10-14, 2026-10-29 and 2026-11-04: A short history of Ollama (llama.cpp created 2023-03-10; GGUF merged 2023-08-21; Ollama v0.0.1 on 2023-07-08; the doorman served by Ollama), A short history of vLLM ("deployed at Chatbot Arena and Vicuna Demo for the past two months", in vLLM's post of 2023-06-20), A short history of Mistral (Lample and Lacroix on the LLaMA paper; "open weights at almost every size"), A short history of Qwen (published 2026-10-23; its licences by size), the seat trials and The August arrivals (the llama3.3:70b row), Two Hours at 12 tok/s — On Battery (the 2.7 tok/s, 41 GiB and 37.6 s figures), the llms.txt reading (the per-arm table), the voice trials (the llama3.1:8b row), the licence ledger (no Llama row; the narrator, gemma4:12b, an Ollama tag read with ollama show), Uncensored models (the "Llama 3 Acceptable Use Policy" read 2026-10-02), The compressed photograph

Corrections and later measurements will be added below, each dated (UTC), with a window at both ends where one applies and saying in plain words what it counts.