Muse Spark

Muse Spark is a natively multimodal foundation model developed by Meta Superintelligence Labs (MSL), a specialized research unit led by former Scale AI CEO Alexandr Wang. Released as the successor to the Llama 4 series, Muse Spark represents a fundamental shift in Meta's AI strategy, moving from an open-weight community focus to a proprietary, product-first approach. The model is built on a completely rebuilt training stack designed to achieve frontier-level performance with significantly higher compute efficiency than previous generations.

The model introduces a novel "thought compression" technique developed during reinforcement learning, where the system is incentivized to produce correct answers while minimizing reasoning tokens. This allows Muse Spark to solve complex problems using an order of magnitude less compute than its predecessors while maintaining high accuracy. It features three distinct operating modes: Instant for rapid conversational responses, Thinking for step-by-step reasoning on difficult tasks, and Contemplating, a multi-agent orchestration mode that runs multiple sub-agents in parallel to tackle demanding scientific and mathematical questions.

Natively multimodal, Muse Spark processes text, image, and audio inputs within a single architecture. It is particularly optimized for Visual Chain-of-Thought, enabling it to reason through image-based problems—such as identifying mechanical parts for troubleshooting or analyzing charts—before producing an output. Additionally, the model has a specialized focus on healthcare; it was trained on data curated in collaboration with over 1,000 physicians to provide factual, comprehensive responses to medical and wellness queries, including the ability to generate interactive health visualizations.

Muse Spark currently powers the Meta AI assistant across the company's ecosystem, including integrated features in Facebook, Instagram, WhatsApp, and the Ray-Ban Meta smart glasses. While the model is initially available as a closed-source product through a private API preview for select partners, Meta has indicated a potential for future open-weight releases in the Muse series.