Meta has rolled out Muse Spark, the first model in its new Muse family developed inside Meta Superintelligence Labs. The release comes after the company rebuilt its entire AI training pipeline from the ground up over the past nine months, complete with fresh infrastructure, architecture, and data systems.
Muse Spark stands out as a natively multimodal reasoning model. It handles text and images together, supports tool use, visual chain of thought, and multi-agent orchestration. Users can already try it on meta.ai and inside the Meta AI app, with a private API preview opening to select partners. A special Contemplating mode that spins up multiple agents to reason in parallel is rolling out gradually.
Stronger Performance Across Key Areas 🤖
In testing, Muse Spark shows competitive scores on multimodal perception, reasoning, health-related questions, and agentic tasks. The Contemplating mode helps it reach 58 percent on Humanity’s Last Exam and 38 percent on FrontierScience Research. Meta also worked with more than 1,000 physicians to shape its health knowledge, so responses stay factual and useful.
Practical examples include turning a photo into a playable Sudoku game, creating interactive tutorials that highlight appliance parts, or analyzing food images to mark recommended items with personalized health scores for users with specific diets. It can even review yoga poses from images, rate form, and suggest fixes.
Efficient Scaling on Three Fronts 💡
Meta scaled Muse Spark along three axes: pretraining, reinforcement learning, and test-time reasoning. The rebuilt pretraining stack delivers the same performance as its earlier Llama 4 Maverick but with over an order of magnitude less compute. Reinforcement learning produces steady, predictable gains without sacrificing diversity in answers. Test-time improvements, including thinking-time penalties and parallel agents, keep latency low while boosting accuracy on tough problems.
Safety checks under Meta’s updated Advanced AI Scaling Framework confirmed strong refusals on high-risk topics such as biological and chemical weapons. The model lacks dangerous autonomous capabilities in cybersecurity or loss-of-control scenarios. Independent review by Apollo Research noted high evaluation awareness but found no issues that blocked deployment.
What Comes Next for Meta AI 📈
Muse Spark powers an updated Meta AI experience today and will expand to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks. Larger models in the Muse series are already in development, following a deliberate, step-by-step scaling approach.
This launch represents a clear shift for Meta, moving away from its previous open-weight Llama strategy toward a more closed model focused on consumer products and personal superintelligence goals. Analysts have noted the progress after a year of heavy investment in talent and infrastructure.
Developers interested in the broader ecosystem can explore related work on Llama models or dive deeper into multimodal AI tools at Meta AI Research. For those building health-focused applications, Meta’s approach to physician-curated data offers a useful reference via AI at Meta. Additional context on frontier model safety appears in coverage from Reuters and performance comparisons in The New York Times.
