The smartest AI Meta: Muse Spark opens a new era

Meta has introduced a new AI model Muse Spark, which is the first step in the development of the Muse line and part of the company’s updated AI strategy. The development belongs to Meta Superintelligence Labs and is already available to users via meta.ai and the Meta AI app, while the API is open in a limited preview format.

Muse Spark is positioned as a natively multimodal model with support for complex reasoning, tooling, and multi-agent interaction. One of the key features is the Contemplating mode, in which several agents process tasks in parallel. This allows us to reach a level comparable to the advanced modes of competitors, including Gemini Deep Think and GPT Pro.

In tests, the model demonstrated 58% in Humanity’s Last Exam and 38% in FrontierScience Research. The company emphasizes that Muse Spark is the basis for creating a personal superintelligence capable of analyzing the user’s environment and helping in everyday tasks, including healthcare. To improve the accuracy of answers, the model was trained with the involvement of more than 1000 doctors.

Meta also reported a significant breakthrough in training efficiency. Thanks to an updated approach to pre-training, the new model achieves a similar level of capability with more than 10 times less computational cost compared to Llama 4 Maverick. Reinforcement learning ensured a steady increase in quality, and optimized reasoning reduced resource consumption without sacrificing accuracy.

Special attention is paid to security. According to the test results, Muse Spark demonstrates the correct rejection of unsafe requests and shows no signs of autonomous risky behavior. Meta notes that the model meets the established criteria for safe deployment and will become the basis for further development of more powerful AI systems.