The exemption doesn't travel — open weights, fine-tuning and the line where you become the provider

Meta's Muse Glimmer, released 10 August 2026 under Apache 2.0, is read as freedom from the AI Act. Article 53(2) lifts only 53(1)(a) and (b), and only for the provider that releases the model that way. The clause that moves an engineer is the Commission's downstream-modifier criterion: a modification whose training compute exceeds a third of the original model's makes the modifier the provider of a general-purpose AI model, with a fallback of a third of 10^23 FLOP where the original figure is unknown. Meta does not publish Muse Glimmer's training compute — because 53(2) lifts the Annex XI documentation in which it would appear — so the fixed fallback governs, and a fixed bar is proportionally far easier to cross against a small open base than against a frontier model. The exemption attaches to how a model is released, not to what it descends from.

Post

Meta shipped a 30-billion-parameter agent model on 10 August under Apache 2.0. Everyone read the licence as an exemption. It is closer to a transfer.

Muse Glimmer is dense, tuned for tool use and long tasks. Meta's own post puts full precision above 55 GB and the 4-bit build under 20 GB — the card already in your workstation.

Now the European part.

Article 53(2) lifts two duties for a model released under a free and open-source licence with public weights: 53(1)(a), the Annex XI technical documentation, and 53(1)(b), the Annex XII pack owed to you downstream. The copyright policy at (c) and the training-data summary at (d) survive it.

Here is the clause nobody quotes.

The Commission's guidelines of 18 July 2025 set an indicative criterion for when a downstream modifier becomes the provider of a general-purpose AI model: the modification's training compute exceeds a third of the original model's. Where the original figure is unknown, the fallback is a third of 10^23 FLOP.

Meta's post does not state Muse Glimmer's training compute. There is no denominator, so the fallback applies — and the reason it is missing is the exemption itself. Annex XI Section 1 point 2(c) is "the computational resources used to train the model".

Rough 6ND arithmetic — mine, not theirs — puts a third of 10^23 FLOP near 180 billion tokens of continued pre-training on a 30B dense model. A LoRA never gets close; a serious continued-pretraining run is in the neighbourhood. Against a frontier base, a third of its own compute sits a hundred times higher. The fixed fallback bites hardest exactly where small open models get modified.

And the exemption does not travel. It attaches to how a model is released, not to what it descends from. Ship your fine-tune with private weights and 53(1)(a) and (b) come back — owed by you, to your users.

What I'd do Monday:

  1. Instrument training runs to emit a FLOP count. You cannot argue a threshold you never measured.
  2. Ask your base provider for its training compute in writing.
  3. Treat the release licence as an architecture decision.
  4. Write the Annex XII pack for your own fine-tune before a customer asks.

Open weights moved the duty. They did not delete it.

Which of your fine-tunes has a measured FLOP count on file this morning?

#AgenticAI #EUAIAct #AIEngineering #OpenWeights #Berlin

First comment

Sources, primary first —

Meta AI Research, 10 August 2026 — the model itself: 30B dense, Apache 2.0, the memory figures and the tool-calling framing. Date and specs taken from Meta's own post, not from coverage. https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model

The model listing, if you want to pull the weights rather than read about them https://developer.meta.com/ai/models/muse-glimmer/

European Commission — the guidelines on the scope of obligations for providers of general-purpose AI models, adopted 18 July 2025. The downstream-modifier criterion lives here. https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers

Commission Q&A on general-purpose AI models — the indicative criterion in its own words: "the training compute used for the modification is greater than a third of the training compute of the original model" https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers

AI Act Article 53 — paragraph 1(a) to (d) and the paragraph 2 open-source carve-out https://artificialintelligenceact.eu/article/53/

Annex XI — Section 1, point 2(c), "the computational resources used to train the model (e.g. number of floating point operations)" https://artificialintelligenceact.eu/annex/11/

Freshfields on the guidelines — useful for the FLOP thresholds side by side, including the fallback figures https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-27-guidelines-on-the-scope-of-obligations-for-general-purpos-102kxqo

One point cut for length: if your modification pushes the model over 10^25 FLOP cumulative, the limitation to "the modification only" falls away and you own the whole model, systemic-risk obligations included. That is a different post.

Claims

Sources

Corrections

None.

AI disclosure

The video for this edition contains AI-generated media: the narration is a synthetic voice and the cover art is generated. The script, the claims and the source verification are human-authored and human-checked. Assets carry IPTC DigitalSourceType and C2PA Content Credentials.