Meta releases Muse Spark Open weights, keeping frontier AI models closed
Meta publishes a coding-focused AI model while stating it will not open-source its most powerful systems due to safety concerns.
Key points
- Meta published weights for Muse Spark Open on Hugging Face, Amazon, and Google Cloud.
- Chief AI Officer Alexandr Wang stated the company will generally keep its most capable models closed.
- The release aligns with the EU AI Act’s exemption for genuinely open models that lack systemic risk.
Meta has published the model weights for Muse Spark Open, a coding and agent-focused artificial intelligence system. According to Bloomberg, the release occurred on Thursday. The company explicitly states that this new model is less advanced than Muse Wonder, a frontier model it launched on Monday.
Muse Spark Open is available on Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden. Nebius, a Dutch firm backed by Meta with up to $27 billion in committed funding, is one of the platforms serving the model. This marks the second open release in the Muse line, following Glimmer in August.
Policy on Open Source and Regulation
Chief AI Officer Alexandr Wang said Meta will generally not open-source its most powerful models. He cited safety and capabilities as the primary reasons for this decision. Wang emphasized that the company intends to continue releasing high-quality open models, but those will not include its top-tier frontier systems like Muse Wonder.
This strategy intersects with the European Union’s AI Act. Article 53 of the legislation exempts genuinely free and open models from certain documentation requirements owed to the AI Office. However, this exemption does not apply to models deemed to pose systemic risk, which must follow all obligations regardless of their license.
Meta had not decided in September whether to publish weights for Muse Spark 1.3. Instead, the company released Muse Spark Open. Glimmer, released in August under the Apache 2.0 license, was a 30-billion-parameter model distilled from Muse Spark. Meta also collaborated with Apple to ensure the new model runs on M5 Ultra Mac Studios.
Leadership Statements and Industry Context
Meta’s founder Mark Zuckerberg told the Sources podcast five weeks ago that it was dangerous for rivals to train advanced models without releasing them. He did not name any specific company in that statement. This view contrasts with the current policy articulated by Wang regarding Meta’s own frontier models.
The European Commission can fine companies up to 3% of global turnover and demand evaluations before release. Meta declined to sign a code of practice designed to operationalize these duties. Last year, Joel Kaplan stated that Europe was on the wrong path regarding these regulations.
Wang said most of the world’s AI will run on open models and argued they should be American. Eight days prior to Wang’s comments, Alibaba chairman Joe Tsai told a conference in Turin that open source was Europe’s only path to technological independence. Meta published no benchmarks for Muse Wonder when it announced the frontier status on Monday.
What should you do?
Security and technology professionals should monitor how organizations integrate open-source AI models into their development pipelines. The distinction between open and closed models affects transparency and potential risk assessments. Teams should review their internal policies on using third-party AI weights, especially those hosted on major cloud platforms.
What to do and how to stay safe: Meta
- Review internal guidelines for integrating open-source AI models into software development workflows.
- Assess the provenance and licensing terms of any new model weights downloaded from platforms like Hugging Face.
- Monitor regulatory updates regarding the EU AI Act and its definition of systemic risk for AI systems.
- Verify that deployed models align with organizational safety standards before allowing access to sensitive code repositories.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Is Muse Spark Open the same as Muse Wonder?
No, Meta states that Muse Spark Open is less capable than Muse Wonder, which remains a closed frontier model.
Where can developers access the Muse Spark Open weights?
The weights are available on Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden.
Why is Meta keeping its most powerful models closed?
Chief AI Officer Alexandr Wang cited safety and capabilities as reasons for generally not open-sourcing the most powerful models.
