Meta releases Muse Wonder AI model, claims parity with ChatGPT and Claude
Meta’s new Muse Wonder model matches competitors in coding and professional tasks, though weight release status remains uncertain.
Key points
- Meta released Muse Wonder, internally known as Watermelon, claiming performance parity with ChatGPT and Claude.
- The company has not confirmed whether it will publish the model’s weights, affecting EU AI Act compliance.
- Chief AI Officer Alexandr Wang stated developers can access the model via the Meta Model API and Muse Code.
Meta has released Muse Wonder, its most capable large language model to date, claiming it matches the performance of ChatGPT and Claude in coding and professional tasks. The model, known internally as Watermelon, marks a significant milestone for the social media giant’s artificial intelligence division. Chief AI Officer Alexandr Wang told Bloomberg that the system achieves parity with leading competitors on financial and legal work.
The release coincides with Meta’s internal target date for the project, which The Information reported in August was set for October. Wang made similar claims about the earlier Muse Spark 1.3 in September, while Mark Zuckerberg previously called the first Muse Spark a world-class assistant in April. However, Meta has not published benchmarks for Muse Wonder to substantiate these performance assertions.
Performance and Infrastructure Context
Meta’s initial Muse Spark model came with a scorecard that placed it fourth on one index, scoring 52 against 57 for leaders. It trailed on two reasoning benchmarks according to Meta’s own figures. Wang argues that benchmarks are weak proofs that can be gamed, urging users to test the model themselves. The spending behind this effort is substantial, with Zuckerberg committing hundreds of billions to infrastructure and paying $14 billion for a stake in Scale AI.
Regulatory and Access Implications
Developers can currently access Muse Wonder through the Meta Model API and Muse Code, and it will power the paid tier of the Muse agent launched last month. A critical regulatory question remains whether Meta will publish the model’s weights. Under Article 53 of the EU AI Act, genuinely open models with public weights and architecture escape certain documentation requirements owed to the AI Office.
Meta has not stated whether Muse Wonder will be open. The company had not decided for Muse Spark 1.3 either, with Zuckerberg signaling caution about opening powerful systems. Meta previously released a distilled 30B model openly but kept its flagship closed. Models with systemic risk owe full obligations regardless of license, and Meta declined to sign the operationalizing code of practice.
What should you do?
Security and technology professionals should monitor Meta’s official announcements for details on weight release and licensing terms. Developers using the Meta Model API should review the specific usage restrictions and data handling policies for Muse Wonder. Organizations relying on AI for coding or professional work should independently test the model’s outputs for accuracy and bias, as claimed parity lacks published benchmarks. Watch for updates on how Meta addresses EU AI Act obligations for systemic risk models.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Meta
- Monitor official vendor announcements for clarification on whether model weights and architecture will be made publicly available.
- Review the specific terms of service and usage restrictions associated with the Meta Model API and Muse Code access.
- Conduct independent testing of AI-generated code and professional documents to verify accuracy, as performance claims are not backed by published benchmarks.
- Assess how the model’s potential classification under the EU AI Act might impact your organization’s compliance requirements and liability.
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
Will Meta publish the weights for the Muse Wonder model?
Meta has not yet confirmed whether it will publish the weights for Muse Wonder, leaving its status under the EU AI Act unclear.
How does Muse Wonder compare to ChatGPT and Claude?
Meta claims Muse Wonder matches ChatGPT and Claude on coding and professional work, though it has not provided published benchmarks to verify this.
How can developers access Muse Wonder?
Developers can access Muse Wonder through the Meta Model API and Muse Code, and it will power the paid tier of the Muse agent.
