Meta Launches Muse Spark, Beats Top AI Rivals

Meta has unveiled Muse Spark, a multimodal reasoning model that outperforms Claude 4.6 Opus, Gemini 3.1 Pro, and GPT 5.4 on several benchmarks, including HealthBench Hard for medical questions and CharXiv Reasoning for scientific chart analysis. The model was developed with a clinical dataset built alongside over 1,000 physicians and requires an order of magnitude less compute than its predecessor, Llama 4 Maverick. Muse Spark is rolling out to Meta AI users over the coming weeks and is available to developers via a private API preview. A Contemplating mode, which deploys parallel AI agents to break tasks into substeps, boosted its score on the difficult HLE benchmark by around 8%. Meta described Muse Spark as the first in a planned series of multimodal reasoning models.