Meta keeps top AI closed, releasing only weaker Muse Spark for coding and agent tasks

Meta published the weights for Muse Spark Open, a coding and agent model described as less advanced than its frontier offering. The release occurred after Bloomberg reported the news on Thursday, following the Monday launch of the closed Muse Wonder model.

The company explicitly stated it will not open-source its most powerful models due to safety and capability concerns. This policy distinction aligns with the EU AI Act’s open-source exemption, which applies to genuinely free models but not to those posing systemic risk.

Strategic context

Chief AI officer Alexandr Wang said Meta intends to keep releasing good open models while withholding its best ones. He argued that most of the world’s AI will run on open models and that they should be American. This stance contrasts with comments from Meta’s founder five weeks ago, who warned that rivals withholding advanced models was dangerous.

Regulatory and supplier landscape

The weights are available on Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden. Nebius, a Dutch firm receiving up to $27 billion in Meta commitments, is also serving the model. Meta previously released Glimmer in August under Apache 2.0 and worked with Apple for M5 Ultra compatibility.

Defender's checklist

Organizations should verify the license terms for any open-source AI models integrated into their workflows. Teams should assess whether the model’s capabilities match their security requirements and monitor for updates to the model’s weights or documentation.

Operational impact

Teams usually evaluate new open-source tools against internal security policies before deployment. Check whether the model’s license permits commercial use and data retention. Inform legal and compliance staff about the distinction between open and closed models to ensure alignment with regional regulations like the EU AI Act.

Background: Data retention policies

Document retention policies for nonprofits define how long you keep records and when you delete them. This reduces the amount of sensitive data exposed during a breach. It lowers storage costs and ensures you meet legal obligations. You must define clear rules for creation, storage, and secure destruction of all organizational records.

Read the full guide: Document retention policies for nonprofits: reduce risk and save space

What to do and how to stay safe: Meta

• Review the Apache 2.0 or specific license terms for Muse Spark Open before integration.
• Verify that the model’s capabilities are appropriate for your specific coding or agent tasks.
• Monitor official channels for updates to the model’s weights or security advisories.
• Ensure your internal AI usage policy accounts for the distinction between open and closed models.

Step-by-step guide: Implementing AI Usage Policies at Work: A Step-by-Step Framework

General security guidance from the InfoSec Ledger newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Is the Muse Wonder model open source?

No, Meta stated that Muse Wonder remains closed and is considered its frontier model.

Where can I access the Muse Spark Open weights?

The weights are available on Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden.

Why is Meta not releasing its most powerful models?

Chief AI officer Alexandr Wang cited safety and capabilities as reasons for keeping top models closed.