Apache 2.0 and Still Not Open Source: What 'Open Weights' Actually Means

On 10 August, Meta released Muse Glimmer, a 30-billion-parameter agentic model, under the Apache License 2.0 with ungated weights on Hugging Face (Meta AI Research; VentureBeat). Corroborated Apache 2.0 is about as permissive as licensing gets: use it commercially, modify it, redistribute it, sue-proof patent grant included, no revenue threshold, no acceptable-use policy bolted on.

By the Open Source Initiative's definition, Muse Glimmer is still not open source AI.

That is not a gotcha. Glimmer is distilled from Muse Spark — Meta's closed, API-only frontier model, launched 8 April 2026 and metered at $1.25 per million input tokens. The teacher model's weights are unpublished, its training data undescribed. A skilled engineer handed everything Meta published about Glimmer could not rebuild Glimmer, because the thing it was distilled from is a black box.

So the licence says "do whatever you want" and the system says "you may not understand me." Both are true simultaneously. This is the single most useful thing to internalise about open weights, and it explains almost every argument happening in this space right now.

Three promises wearing one word

When somebody says a model is "open," they are making one of three claims, and they are usually not saying which.

Open weights means the numbers are downloadable. That is a statement about artefact availability, and nothing else. It does not tell you what you may legally do with them, whether you can inspect how they came to be, or whether you can reproduce them. It is a distribution fact.

Open source — in the sense the term has meant since 1998 — means you receive the freedoms to use, study, modify and share. That is a statement about rights. Crucially, "study" and "modify" carry an implicit requirement: you need enough of the preferred form of the work to actually exercise them. In software that means source code. In a model, the analogous question is genuinely hard, and it is the entire fault line.

Open data means the training corpus is published. That is a statement about provenance. It is the rarest of the three by an enormous margin, and it is the one that determines whether "study" means anything at all.

A model can have any combination of these. Muse Glimmer has the first two and not the third. Llama has the first, a heavily conditioned version of the second, and not the third. GPT-class API models have none. OLMo has all three.

Confusion between these is not merely semantic. It has begun to carry legal weight, which is where the last section of this piece ends up.

The licence gradient

Downloadable weights arrive under wildly different terms. Four tiers, roughly, from most to least permissive.

Tier 1 — Actual open source licences

Apache 2.0 and MIT, applied to the weights. No use restrictions, no user thresholds, no downstream obligations beyond attribution and notice.

This tier has grown substantially. Qwen moved to Apache 2.0 from Qwen3 onward, DeepSeek shifted to MIT-licensed weights with V3 and R1, the Mistral 3 family launched directly under Apache 2.0, and gpt-oss arrived under Apache 2.0 as well. Muse Glimmer now joins them. Corroborated

One caveat worth carrying: a vendor being "an Apache 2.0 vendor" is not a stable property. Alibaba also ships certain Qwen releases under a source-available licence and a non-commercial research licence. The licence attaches to the release, not the brand. Check per model, every time.

Tier 2 — Community licences with conditions

The Llama Community License is the archetype. Commercial use is permitted, but three conditions ride along:

• A user threshold. Any licensee whose products — expressly including affiliates — exceed 700 million monthly active users must request a separate licence from Meta (Llama 3.1 Community License). Affiliate scope generally reaches entities connected by ownership stakes of 50% or more. The practical hazard is not organic growth; it is acquisition. Get bought by somebody already over the line and your continued use depends on Meta's goodwill, renegotiated at the worst possible moment.
• An acceptable-use policy incorporated by reference, which can be revised by the licensor without your involvement.
• A non-compete on outputs — you may not use outputs to improve a competing foundation model.

None of these are unreasonable business terms. All of them are disqualifying for open source, because open source does not permit discrimination against fields of endeavour or against classes of user. This is why the accurate description of Llama is source-available or, better, open weights.

Tier 3 — Use-restricted licences with pass-through

Gemma is the instructive case, because the obligations are structural rather than merely restrictive. Reading the Gemma Terms of Use directly:

• § 3.2 — "To the maximum extent permitted by law, Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement."
• § 3.1(1) — you must include the use restrictions "as an enforceable provision in any agreement ... governing the use and/or distribution of Gemma or Model Derivatives."
• § 1.1(e) — Model Derivatives include models produced by weight transfer or distillation.

Stack those three and the shape becomes clear. The restrictions survive fine-tuning. They survive distillation. They must be propagated contractually to everyone downstream of you, which means you are conscripted as an enforcement agent for terms you did not write. And a reserved right of remote restriction sits over it all.

Inference For a company embedding a model in a product with its own commercial terms, the pass-through obligation is usually a bigger practical problem than the prohibited-use list itself. The prohibited uses are things most businesses were not going to do. The pass-through is a clause you must successfully negotiate into every customer contract, forever.

Tier 4 — Research-only and non-commercial

Weights you can download and legally cannot ship. Perfectly legitimate as a release mode; frequently misread as open because the download button looks identical. The download button always looks identical. That is the core problem.

What the OSI definition actually asks for

The Open Source Initiative published version 1.0 of the Open Source AI Definition (OSAID) requiring four freedoms — use, study, modify, share — applied to an AI system, plus disclosure across three components: weights, the code used to train and run the system, and Data Information.

Data Information is where the argument lives. The requirement, verbatim:

"Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system."

That is deliberately not "publish the training data." It asks for complete descriptions of provenance, scope, characteristics and labelling procedures, plus listings of publicly available and third-party sources — including sources available only for a fee.

The compromise satisfied nobody completely. Critics — the Software Freedom Conservancy's position has been among the most pointed — argue that a definition permitting a description in place of the data guts the freedom to study, since you cannot audit a corpus you cannot see, and cannot reproduce a system from prose. Defenders note that a definition requiring full data publication would be met by essentially nothing, including models everyone intuitively considers open, given that most large corpora contain material the trainer has no right to redistribute.

Two procedural criticisms attach as well: OSAID 1.0 was approved by the OSI's board rather than its membership, and it was drafted fast. The OSI has positioned 1.0 as a starting point with a revision process running through Q4 2026. Unverified — the schedule comes from OSI's stated roadmap; whether 1.1 or 2.0 actually lands in that window is not yet established.

The practical result today: OLMo, from the Allen Institute for AI, is the only major family that clears the strictest reading — Apache 2.0 weights, the Dolma training corpus, training code, evaluation suite, intermediate checkpoints and training logs, all published (Ai2). Corroborated OLMo 3 shipped in November 2025 and closed much of the gap to comparable open-weight peers, though not to the frontier.

That is the honest state of play. One family is open source AI under the definition. Everything else on your shortlist is open weights.

Why the argument has teeth

For years this was a taxonomy dispute among people who care about taxonomy. The European Union's Artificial Intelligence Act (EU AI Act) changed that by attaching consequences to the category.

General-purpose AI (GPAI) models released under a free and open-source licence receive an exemption from parts of the GPAI obligations. Qualifying requires three conditions together:

1. Release under a free and open-source licence permitting access, use, modification and distribution;
2. Public availability of the parameters — weights, architecture information, and usage information;
3. The model is not provided against a price or otherwise monetised.

Three details decide real cases.

The exemption is partial, not total. It removes the obligation to provide documentation to downstream providers and to authorities on request. It does not remove the training-data summary requirement or the copyright-policy requirement. Those apply regardless.

Monetisation is fatal to it. A model behind any paid arrangement loses the exemption even if the weights are downloadable under a permissive licence. Inference This is where dual-track strategies — free weights plus a paid hosted tier of the same model — get genuinely uncertain, and it is the question I would want counsel on before relying on the exemption commercially.

Systemic risk overrides it. A model classified as carrying systemic risk does not get the open-source exemption at all.

And the exemption is about the provider's obligations. Deploying an open model in a high-risk context leaves your deployer obligations completely intact. "We used an open model" is not a compliance argument.

Reading a licence in five minutes

A practical sequence, ordered by how often each one actually bites:

1. Name the licence exactly, including version, for the specific release. Not the vendor's reputation. The file.
2. Search for a user or revenue threshold. If one exists, model your acquisition scenarios, not just your growth curve.
3. Find every document incorporated by reference — acceptable-use policies, prohibited-use policies. Check whether the licensor may revise them unilaterally.
4. Check whether obligations pass through to your customers. This is the clause most likely to collide with contracts you have already signed.
5. Check the definition of derivative works. If distillation and weight transfer are captured, you cannot launder the terms off by fine-tuning.
6. Check what is said about training data — and treat "trained on a diverse mixture of publicly available sources" as what it is, which is nothing.

The part worth remembering

Meta went from a community licence, to a closed metered API, to Apache 2.0, inside sixteen months. The direction of travel was not principled in either direction; it tracked competitive position. Openness is a dial that vendors turn, not a property that projects possess.

Which is exactly why the vocabulary matters. If "open" collapses into a single approving adjective, you lose the ability to ask the three questions that actually determine what you can do: May I use this? — the licence. Can I understand this? — the data and code. Can I get this without permission? — the weights.

Muse Glimmer answers yes, no, yes. Llama answers mostly, no, yes. OLMo answers yes, yes, yes. GPT-class models answer no, no, no.

Four different situations. One word doing all the work, and doing it badly.