Meta Introduces 'Muse Spark' Next-Gen AI With Contemplating Mode

Meta has introduced Muse Spark, the first model from its Meta Superintelligence Labs. It's described as the result of a broader overhaul of the company's AI stack, with the system built as a natively multimodal model that integrates visual understanding, tool use, and multi-agent orchestration into a single architecture.



Meta worked with more than 1,000 physicians to refine its training data. On the "HealthBench Hard" test, Muse Spark scored 42.8, more than double the 20.6 posted by Google's Gemini 3.1 Pro (High). The model can generate interactive charts that explain everything from food nutrition to which muscles are engaged during exercise, with capabilities similar to platforms like ChatGPT Health and Perplexity Health.


To handle more complex tasks, Meta is introducing "Contemplating mode." The feature runs multiple agents in parallel to reason through problems before generating a response. On the FrontierScience Research benchmark, Muse Spark achieved 38.3 percent accuracy, ahead of both GPT 5.4 Pro at 36.7 and Gemini 3.1 Deep Think at 23.3.



The model also shows strong performance in multimodal reasoning. On the CharXiv Reasoning benchmark, Muse Spark scored 86.4, outperforming several competing models on figure understanding tasks. Meta says these capabilities enable use cases like interactive troubleshooting, where users can point a camera at a home appliance and receive annotated guidance.

The model is significantly more efficient than its previous Llama 4 Maverick, requiring over an order of magnitude less compute to reach similar performance levels. However, its advanced capabilities have also prompted additional safety analysis. Third-party testing by Apollo Research found that Muse Spark exhibits a high level of "evaluation awareness." The model frequently identified when it was in a testing environment and reasoned that it should behave honestly because it was being monitored, which Meta says it is continuing to study.

The launch comes as Apple prepares its own significant software transition. While Apple prepares a rebuilt Siri for iOS 27 that will rely on Google's models, Meta is investing across its own stack, including infrastructure like the Hyperion data center. This stands in contrast to the capital-light strategy Apple has used to avoid multibillion-dollar server investments.


Muse Spark is available starting today at meta.ai and through the Meta AI mobile app, with a private API preview now open for developers and enterprise partners.