Apache 2.0, and a second file that rules out defence and health

If your team downloaded an open-weight model tomorrow for a project on a plant, in a hospital, or for defence, who in your organisation would say yes or no — and which document would they read it on? The licence label on the repository card, or the second file sitting in the same folder, the one almost nobody opens?

This time there is no deception to report, and that needs saying upfront. In recent days we have written more than once about licences declared incorrectly or incompletely. Here it is the opposite: the publisher is verified, the declared licence is genuine, and the file containing it really does carry the text of the Apache License 2.0. The problem, this time, is not the licence. It is what sits beside it.

The model, and who publishes it

meta-models/Muse-Glimmer-30B appeared on Hugging Face on 10 August 2026 — created on the 9th, last card update at 08:15:02 UTC on the 10th, per Hugging Face’s API at the time of our query. At that moment it registered 675 likes and 0 downloads: an API figure on a model that had just launched, not a conclusion. Its trajectory score (660) still placed it third in the trending ranking, behind only MiniMax-H3 and DeepSeek-V4-Flash.

Before reading any licence, the first check concerns who publishes. The organisation meta-models shows, via the API, a fullname of “Meta Inc.”, isVerified: true, 150 users, 703 followers and four models published between 9 and 10 August — Muse-Glimmer-30B, -GGUF, -assistant, -ExecuTorch-PTE. The handle is not the one historically used for the Llama models, which is exactly why you check the verification flag rather than the name in the URL: here it clears the bar. The card attributes the work to “Meta Superintelligence Lab”, “Model Release Date: August 2026”, and describes a model of ~29.6 billion parameters, built to run entirely locally, “without requiring cloud infrastructure or network access”.

Two documents, only one indexed

The card’s frontmatter declares license: apache-2.0; the body repeats it: “License: Apache 2.0”, and further down “All artifacts are released under Apache 2.0”. The LICENSE file — 11,358 bytes — really does contain the text of the Apache License 2.0, checked against the original: no substitution. On this point, which we always check first, the repository is in order.

The repository, however, also contains a second file, USAGE_POLICY.md — 5,230 bytes —, which opens like this, verbatim: “Meta is committed to promoting safe and fair use of its tools and features, including Muse Glimmer. This Usage Policy (“Policy”) applies to your access or use of Muse Glimmer.” Just below: “Muse Glimmer is not intended for individuals under the age of 18”. Among the prohibited uses, the ones that concern exactly this article’s readers:

“Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) […] or to the U.S. Biological Weapons Anti-Terrorism Act of 1989 or the Chemical Weapons Convention Implementation Act of 1997”;
“Operation of critical infrastructure, transportation technologies, or heavy machinery”;
“Engage in the unauthorized or unlicensed practice of any profession including […] financial, legal, medical/health, or related professional practices”;
“Collect, process, disclose, generate, or infer private or sensitive information about individuals, including […] identity, health, or demographic information […]”.

Among the obligations in the same file there is also: “Fail to appropriately disclose to end users any known dangers of your AI system” — listed among the things not to do.

Where the tension sits, and where we don’t give a legal opinion

The Apache 2.0 is permissive: by design it contains no restriction on field of use. Defence, health, critical infrastructure — it says nothing about any of them. The card, under “Out-of-scope”, states: “Use in any other way that is prohibited by the Apache 2.0 License terms”. But the bans on defence, critical infrastructure and health do not sit in the Apache 2.0: they sit in the second file, which the card cites only once, in the “Responsible Use” paragraph: “Our Usage Policy can be found here [link]”.

We do not give a legal opinion: whether that policy contractually binds someone downloading Apache 2.0 weights is a question for the reader’s lawyer, not an article. What we can demonstrate, checked file by file, is simpler: an inventory built on the licence label — the license: apache-2.0 field any scraper reads in a second — never sees that second file. Not because it is hidden: it sits in the same folder, listed alongside LICENSE and config.json. But no census pipeline that stops at the licence field opens it.

See the service · Talk to an engineer

The technical details that matter to whoever decides

The weights come in two safetensors files of 49.95 GB and 9.60 GB — around 59.6 GB in total, in bfloat16. The architecture is MuseGlimmerForConditionalGeneration, model_type muse_glimmer: its own, not a variant your inference engine already recognises. transformers_version is 5.15.0.dev0 — a development version, not a stable one: it weighs as much as the licence. Text side: 52 layers, max_position_embeddings 131072, vocab_size 202048, num_key_value_heads 2. Vision side: 50 layers, patch_size 14. The model is multimodal (pipeline_tag image-text-to-text). GGUF conversions already exist, from the organisation and from third parties (unsloth/Muse-Glimmer-30B-GGUF, trending the same day): a conversion redistributes the weights, not the Usage Policy — the same point behind the derivatives carrying a different licence down the chain.

Why it matters even when everything is in order

The MiniMax H3 and Qwen3-8 cases each had a flaw to flag: an incorrect or missing licence. Here the publisher is verified and the licence genuine — and that is what makes the case more instructive, because it shows that even when everything is in order, the label is not the terms. The field an inventory reads automatically does not contain the sentence that changes the answer for a hospital, a utility or a defence office: it sits in a file alongside it, under a different name, that no pipeline opens alone.

From 2 August 2026, Article 101 of Regulation (EU) 2024/1689 is also in play: the Commission can fine providers of general-purpose models directly, up to 3 per cent of total worldwide annual turnover or 15 million euro, whichever is higher. The national penalties under Article 99 had already applied since 2 August 2025. Whoever adopts a model like this in a sensitive context must show, with a dated record, which documents they collected and under what conditions — not merely which licence line appeared on the card.

What to do, in practice

Don’t stop at the license field: always open the repository’s full file tree, looking for policies or restrictions under names other than LICENSE.
If a usage file accompanies the licence, read it in full before deciding on the use case.
For every model under evaluation, record the complete list of documents present, not just the licence.
Repeat the check on every update: a file can change without the license line moving.

The two axes, applied to this case

Complying: the inventory of open-weight models becomes a check that runs on the client’s own systems — for each model, repository and revision (sha), declared licence, which other documents are present and what they say, which use cases it has been approved for and by whom. With the dated record ready to show an inspector or a board.

Deciding: the same system brings the model inventory, contracts, archives, line-of-business systems and documents into a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. When a new model comes out, the answer to “can we use it, for what, and who approved it” arrives in hours. Muse Glimmer is built to run locally: the technical perimeter is already right — it’s the usage decision that still needs putting in writing. Always in two delivery modes: on-premise, on self-contained machines with no deep integration into the client’s network, or dedicated cloud with a data centre in Italy, always with shared management.

Could you say, today, without opening a single repository, which other documents accompany every open-weight model running in your organisation — and for which uses, in defence, in healthcare or on a plant, those documents say no? Talk to one of our engineers: the first session is at no cost, and produces the dated inventory of models in use regardless, with licence, other documents present and approved use cases — including the boxes that stay blank.