Muse Spark is Meta Superintelligence Labs' first Muse-family model, built for native multimodal reasoning, tool use, visual chain of thought, and multi-agent orchestration. Use this page to compare its official demos, benchmark signals, scaling data, and MyClaw workflow fit.
Use Muse Spark when an agent needs to reason over images, tools, health context, and parallel reasoning paths in one hosted workspace.
Muse Spark is designed to integrate visual information across tools and domains, including visual STEM questions, entity recognition, and localization.
Meta highlights visual reasoning traces and dynamic annotations, which help the model explain or troubleshoot what it sees instead of only returning text.
Muse Spark supports tool use and multi-agent orchestration, making it more relevant for agent workflows than a plain chat-only model.
Meta says it worked with over 1,000 physicians to curate training data for more factual and comprehensive health responses.
For difficult reasoning, Contemplating mode coordinates multiple agents in parallel rather than relying on one longer single-agent pass.
Muse Spark is available in Meta AI and the Meta AI app, while private API preview access is limited to select users. MyClaw tracks the model as availability expands.
These demos show the clearest user-facing shape of Muse Spark: visual annotation, health explanation, and movement analysis powered by multimodal reasoning.
Muse Spark can annotate objects and steps in a scene, making it useful for troubleshooting, instruction following, and visual explanations.
Meta shows Muse Spark turning visual context into interactive health and nutrition displays that are easier to inspect than plain text.
Muse Spark can compare movement, mark body positions, and explain how a pose or exercise form can improve.
After the demos, use Meta's benchmark chart as the evidence layer: it reports performance across reasoning, science, multimodal perception, health, and agentic tasks.
Meta frames Muse Spark progress around three levers: more efficient pretraining, smoother reinforcement learning gains, and more efficient test-time reasoning.
Meta says architecture, optimization, and data curation changes helped Muse Spark extract more capability from training compute.
The official scaling-law chart compares Muse Spark with Llama 4 Maverick and leading base models, supporting Meta's claim that the new recipe is significantly more compute efficient.
Meta says its RL stack produces smoother improvements, even though large-scale reinforcement learning can be unstable.
The training plot shows pass@1 and pass@16 rising with RL steps, while the held-out evaluation chart suggests those gains generalize beyond tasks seen during training.
Muse Spark combines thought compression with parallel agent orchestration to spend reasoning budget more efficiently.
Meta says thinking-time penalties can reduce unnecessary reasoning tokens, while multi-agent thinking can improve performance at comparable latency.
Meta evaluated Muse Spark across frontier risk categories, alignment behavior, adversarial robustness, and evaluation awareness before deployment.
Meta reports strong refusal behavior across high-risk biological and chemical domains and says Muse Spark stayed within safe margins for the measured frontier risk categories.
Apollo Research found high evaluation awareness on a near-launch checkpoint. Meta says follow-up work did not make this a blocking release concern, but it remains an area for research.
Meta Muse Spark
Use MyClaw to evaluate Muse Spark-style multimodal workflows with files, tools, browser-ready execution, and a managed agent workspace as model access expands.