Meta Releases Weaker Open AI Model, Keeps Frontier Muse Wonder Closed

Meta Releases Weaker Open AI Model, Keeps Frontier Muse Wonder Closed

Meta has published the weights for Muse Spark Open, a coding and agent model it describes as less advanced than the Muse Wonder frontier model it launched on Monday, according to a report by Bloomberg on Thursday.

The release places Meta's open-source policy on the record: the company will generally not open-source its most powerful models, a distinction that mirrors the split at the heart of the European Union's AI Act open-source exemption.

Not Meta's best model

Muse Spark Open is built for coding and agents, and Meta says it is less capable than Muse Wonder, which the company said on Monday had reached the frontier. No benchmarks were published for Muse Wonder, and that model remains closed.

The weights for Muse Spark Open are available on Hugging Face, Amazon SageMaker AI and Google Cloud Model Garden. One platform serving it is Nebius, the Dutch firm to which Meta has committed up to $27 billion.

Chief AI officer Alexandr Wang said the decision to keep the most capable systems closed was deliberate. "Because of both safety and capabilities, we will generally not open-source our most powerful models," he said. He added that Meta intends to keep releasing good open models.

Second open release from Muse line

The publication is the second open release from the Muse line. In August, Meta published Glimmer, a 30-billion-parameter model distilled from Muse Spark, under the Apache 2.0 licence. It also worked with Apple so the new model runs on two M5 Ultra Mac Studios.

In September, the company had not decided whether to publish the weights for Muse Spark 1.3. It has answered with a different model instead.

The European regulatory context is central to the release. Article 53 of the AI Act exempts a genuinely free and open model from the documentation owed to the AI Office. That exemption stops at systemic risk, where every obligation applies regardless of the licence.

European rules and market pressures

The Commission can fine companies 3% of global turnover and demand evaluations before release. Meta declined to sign the code of practice written to operationalise those duties, and Joel Kaplan said last year that Europe was on the wrong path.

Meta's founder took a different public position five weeks ago. Mark Zuckerberg told the Sources podcast that rivals training more advanced models and then not releasing them was quite dangerous. He named no company.

Wang said most of the world's AI will run on open models, and that they should be American. Eight days earlier, Alibaba chairman Joe Tsai told a Turin conference that open source was Europe's only way to independence. Wang expects Chinese developers to build on Muse Spark Open, and said it might even be good if they adopted its principles.

Europe's own answer is already in the market. Mistral launched Large 4, a trillion-parameter open-weight model, to take on China's lead in this category.

The timing keeps the focus on how far openness will extend as models grow more capable. Meta's approach suggests a durable split between the systems it releases and those it holds back, while European regulators, researchers and companies weigh an exemption designed for genuinely open models against the obligations that apply once systemic risk is in play.

More news

Sitemap