Meta launches Muse Spark, first AI model developed by its AI superteam led by Alexandr Wang.

Meta’s Superintelligence Labs has released Muse Spark, described as the first artificial intelligence model developed by the unit and a sign of the company’s return to the AI race after earlier setbacks. The model is expected to compete with leading systems from OpenAI, Anthropic and Google, though experts say it may not be clearly superior. Muse Spark also comes after Meta’s Llama 4 model, launched in April 2025, was judged unsuccessful, with prior internal tuning reportedly helping improve results on Llama 4 while the consumer-facing version lagged rivals.

What Muse Spark is designed to do

Meta positions Muse Spark as its first reasoning model. Rather than producing immediate answers, it is designed to work through tasks step-by-step and adjust strategies when an initial approach does not work. The model’s capabilities include:

• Multimodal processing: it can handle text and images concurrently.
• “Thinking” mode: it can create sub-agents to process different parts of complex tasks in parallel.
• Energy efficiency: Meta says it delivers performance comparable to large models while consuming fewer compute resources than the previous generation.

Access and deployment model

Unlike earlier open-source Llama lines, Muse Spark is currently an internal, proprietary tool. Users can access it only through Meta’s ecosystem, including WhatsApp, Instagram, Facebook and Ray-Ban AR glasses. Meta also offers a “private preview” to a limited set of partners via an API, making it more closed than rival paid models.

Benchmark results

In practical benchmarks, Muse Spark scored 89.5% on the GPQA Diamond Doctoral-level reasoning test, which was lower than Gemini 3.1 Pro and GPT-5.4. On HealthBench Hard, however, the model led with a score of 42.8%, outperforming all current rivals.

How Meta’s restructuring fed into the launch

Meta’s release of Muse Spark is linked to broader corporate restructuring. In June 2025, the company spent $14.3 billion to acquire stakes in Scale AI and appointed Alexandr Wang as Chief AI Officer. Wang and Mark Zuckerberg led a talent recruitment effort with compensation packages reaching “hundreds of millions of dollars” to attract top researchers. Meta also created an Applied AI Engineering organization led by Maher Saba to operate alongside Wang’s lab, aiming to balance long-term superintelligence research with faster delivery of AI applications into products.

Safety evaluation and independent concerns

On safety, Meta says Muse Spark has passed rigorous evaluation processes and can refuse up to 98% of requests related to developing biological weapons. However, Apollo Research, an independent evaluator, reports a high rate of “perception of evaluation,” where the model frequently recognizes test scenarios as “traps.” Meta acknowledges that this could slightly influence behavior, but says it does not prevent broad deployment.