But Is It Open?The online edition
Type & theme

Make yourself comfortable

Page color
Text size
20 px
Reading width
Ask about this chapter ↗

Your settings and place stay in this browser.

Chapter 12 · Timothy O’Brien

Open Weights Are Not the Whole Story

5 minute read12 source notesFree online edition

Meta’s Llama 4 model card and license are public. They describe a substantial release: downloadable weights, supported uses, performance evaluations, and information about training.1 They also let a reader test an openness claim without accepting or dismissing the whole release at once.

Chapter 11 separated the parts of a model and the process that made it. Here the task is to read what the publisher has actually supplied. A technical report, a license, and a contribution channel answer different questions. Treating one as evidence for all three is how an accurate statement, “the weights are available,” turns into a much broader claim.

First check the permission

Llama 4’s weights are available to eligible users under Meta’s custom Community License, with an Acceptable Use Policy incorporated by reference.2 The terms permit substantial work with the model but limit the grant in ways Apache and MIT don’t.

If the licensee and its affiliates had more than 700 million monthly active users in the month before release, they need a separate license from Meta before exercising the rights. Redistribution carries attribution requirements. The Acceptable Use Policy also excludes EU-domiciled individuals and companies with an EU principal place of business from the grant for multimodal Llama 4 models, while exempting end users of products incorporating them.3

Those are concrete conditions a user can check. The grant and its restrictions don’t depend on whether the announcement calls the release open. Conversely, a custom license doesn’t make the ability to run and modify weights worthless. It means the permitted work has to be described accurately.

The exact release matters. GLM-4-9B-Chat’s 2024 license required commercial registration with Zhipu; GLM-4.5 moved to MIT.45 The later GLM-5.3 license again adds a condition: a qualifying Model-as-a-Service operator above $10 billion in aggregate twelve-month revenue must pass a security review set by Z.AI.6 That is a vendor review gate, not a requirement to obtain government permission. Neither “GLM is MIT” nor a remembered restriction on an older version settles the current checkpoint’s terms.

Chapter 15 examines these clauses in more detail. For now, the useful result is specific: this user may do this work with this release, under these conditions.

Then read the training disclosure

Llama 4’s card describes a mixture of publicly available and licensed data and information from Meta’s products and services. It gives training-token counts, a knowledge cutoff, and environmental estimates.7 Those disclosures let a reader understand the scale and broad categories of the work. They don’t identify the full corpus or supply a complete training recipe.

The gap matters differently depending on the question. Someone testing a model on private documents can measure its responses directly. Someone investigating whether particular material entered training needs provenance information that a benchmark score cannot provide. Someone trying to repeat the training needs processing code, source details, configurations, and enough resources to run them.

It helps to write down the unanswered question rather than treating “transparency” as one quantity. A data-category description doesn’t establish how rights were obtained for every item. A published benchmark doesn’t establish how the model will behave in a different language or workload. A missing corpus doesn’t prevent all behavioral research, but it limits claims about how the artifact was produced.

The OSI’s AI definition asks for enough data information to let a skilled person build a substantially equivalent system, alongside the training and running code and parameters. It accommodates data that cannot itself be shared.8 Full corpus publication is one form of disclosure; it isn’t the definition’s sole test. Llama’s downloadable weights and broad card should therefore be assessed against the actual requirements, not against an invented demand that every release publish every datum.

Look for a way to challenge the result

A model card is the publisher’s account. A reader can inspect its claims, run evaluations, and report failures. Joining the development process is another matter. A public issue tracker doesn’t by itself give outsiders a role in choosing the next training set, changing the release policy, or approving a checkpoint.

This is where open development and open-source terms need separate names. A company can release code under a qualifying license while retaining control over its roadmap. A project can welcome public discussion without granting unrestricted rights to the artifact. Participation tells us who can shape the work; the license tells us what users may do with what has been released.

The questions are observable. Does the project accept outside code or data contributions? Can a contributor follow a proposal through review to a decision? Are decisions and release criteria recorded? Some answers may be plainly yes or no. Others require reading the process and seeing whether it works in practice.

Why disclosure stops short

Publishing more isn’t costless. Training data can contain personal information, confidential material, and works subject to contested rights. The New York Times’s complaint against OpenAI and Microsoft alleges unauthorized use of its articles for training.9 The allegation doesn’t decide the case, but it shows why provenance is a legal question as well as a technical one.

Human labor belongs in the account too. TIME reported that Kenyan contractors doing OpenAI-related safety work took home roughly $1.32 to $2 an hour for reviewing disturbing material.10 A description of “human feedback” can omit the conditions under which it was produced. More informative documentation would let readers examine those conditions without exposing workers’ private information.

There are legitimate reasons to withhold particular records. Publishing private data to prove transparency would create another problem. The EU AI Act requires a sufficiently detailed summary of training content rather than publication of the whole corpus.11 That requirement serves a different purpose from OSI’s definition, so satisfying one should not be presented as automatically satisfying the other.

Ai2’s OLMo releases show a more extensive disclosure approach, including training data, code, recipes, evaluations, and intermediate checkpoints.12 That gives another group more to inspect and reproduce. It doesn’t make the cost of doing so disappear, nor does it mean every lab has the same data rights. It does provide a concrete comparison when a publisher says a thinner release is all that can be made public.

Keep the conclusion as specific as the evidence

After reading a release, a useful assessment should say what the user can obtain and do, what its training documentation supports, and where the publisher retains control. For Llama 4, the weights create real independence from a hosted API; the custom terms restrict that independence for particular users and uses; the training disclosure doesn’t provide a complete reproducible pipeline.

That conclusion preserves the value of the release without letting it stand for more than it supplies. Open source has a definition. Practical openness also includes deployment, provenance, and participation, which deserve their own evidence. The next chapter organizes those checks into the Openness Profile. The field guide that follows applies it to concrete releases, where a publisher’s name is rarely enough to decide the result.

Notes

  1. Meta, “Llama 4 Model Card,” meta-llama/llama-models on GitHub, model release date April 5, 2025, https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md. ↩

  2. Meta, “Llama 4 Community License Agreement,” Llama 4 version effective date April 5, 2025, https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE. Section 1.b.iv incorporates the Llama 4 Acceptable Use Policy by reference. ↩

  3. Meta, “Llama 4 Community License Agreement,” sections 1.b.i (“Built with Llama”) and 2 (700 million monthly active users, “sole discretion”), https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE; Meta, “Llama 4 Acceptable Use Policy,” https://www.llama.com/llama4/use-policy. The EU exclusion is in the Acceptable Use Policy, not the license itself: for multimodal Llama 4 models, the section 1(a) grant does not extend to individuals domiciled in, or companies with a principal place of business in, the European Union, though end users of products built on them are not affected. ↩

  4. THUDM / Zhipu AI, “The glm-4-9b License,” section 2, glm-4-9b-chat on Hugging Face (now under the zai-org account), https://huggingface.co/THUDM/glm-4-9b-chat/blob/main/LICENSE. The registration form is at open.bigmodel.cn. The GLM-4 repository dates the GLM-4-9B release to June 5, 2024: https://github.com/THUDM/GLM-4. ↩

  5. Z.ai, “GLM-4.5,” model card on Hugging Face, https://huggingface.co/zai-org/GLM-4.5, which states the models “are released under the MIT open-source license and can be used commercially.” ↩

  6. Z.AI, “GLM-5.3 License,” 2026, section 2, https://huggingface.co/zai-org/GLM-5.3/raw/main/LICENSE, accessed November 15, 2026. The gate applies if the licensee or affiliates operates a Model-as-a-Service business and their aggregate revenue exceeds $10 billion over any consecutive twelve months. Z.AI reasonably determines the scope and method of the required security review. The condition is imposed by the vendor and does not require government permission. ↩

  7. Meta, “Llama 4 Model Card,” https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md. The card describes the training data as “a mix of publicly available, licensed data and information from Meta’s products and services,” gives token counts (about 40T for Scout, 22T for Maverick), an August 2024 knowledge cutoff, and training greenhouse gas estimates. ↩

  8. Open Source Initiative, “The Open Source AI Definition 1.0,” https://opensource.org/ai/open-source-ai-definition. The definition requires the freedoms to use, study, modify, and share, with data information, code, and parameters as the preferred form for modifications. Data information must support construction of a substantially equivalent system and explicitly accommodates unshareable data; it is not a blanket requirement to redistribute every training datum. ↩

  9. Complaint, The New York Times Company v. Microsoft Corporation et al., No. 1:23-cv-11195 (S.D.N.Y. filed December 27, 2023), paragraph 2, https://www.courthousenews.com/wp-content/uploads/2023/12/new-york-times-microsoft-open-ai-complaint.pdf. These are the plaintiff’s allegations. ↩

  10. Billy Perrigo, “Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic,” TIME, January 2023, https://time.com/6247678/openai-chatgpt-kenya-workers. TIME reports a take-home wage of “between around $1.32 and $2 per hour” for Sama data labelers working for OpenAI; Sama told TIME workers could earn $1.46 to $3.74 an hour after taxes. ↩

  11. Regulation (EU) 2024/1689 (AI Act), Article 53(1)(d), AI Act Service Desk, European Commission, https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53. Providers of general-purpose AI models must publish “a sufficiently detailed summary about the content used for training,” following a template from the AI Office. ↩

  12. Ai2, “Olmo 3: Charting a path through the model flow to lead open-source AI,” November 20, 2025, https://allenai.org/blog/olmo3; Ai2, “OLMo 2,” https://allenai.org/olmo2. ↩

From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.

Chapter 125 min leftNext →

The book, chapter by chapter.

Source note

Go to the notes & sources ↓

The vocabulary.

A short guide to the book’s recurring terms. Follow each chapter for the argument and its sources.

Open weights
A release makes a trained model’s learned parameters available. The accompanying terms decide what you can do with them. Access to weights alone does not provide the training data, code or recipe. Read the chapter →
Copyleft
A way of using license conditions to preserve recipients’ freedoms when covered software is passed on. The obligations depend on the license and on what you distribute or make available. Read the chapter →
Permissive licenses
Licenses such as MIT and Apache 2.0 grant broad reuse rights with comparatively few conditions. Notice obligations still apply, and a software license does not automatically cover a release’s data or other components. Read the chapter →
Reproducibility
The ability to reconstruct and examine how an artifact was produced using the resources that are published. Running an existing checkpoint is a different achievement from rebuilding its training. Read the chapter →
Governance
Who can participate, who makes decisions, and how authority over a project or release is shared. A right to use the code does not by itself create a voice in upstream decisions. Read the chapter →
Training data
The examples used to train a model. Publishing the data, documenting its provenance and describing how it was filtered are distinct kinds of disclosure; privacy and other rights may limit what can be shared. Read the chapter →

A passage worth sharing.

Select a passage in the chapter, or share the one you’re reading.