Meta Platforms Inc. has launched Muse Spark, a new reasoning model demonstrating significant prowess in answering health questions and analyzing multimodal data.
The company plans to integrate Muse Spark into its consumer-facing Meta AI service in the coming weeks. Developers will also gain access via an application programming interface, currently in private preview.
Meta claims Muse Spark surpasses industry benchmarks like Claude 4.6 Opus, Gemini 3.1 Pro, and GPT 5.4, particularly on HealthBench Hard, an evaluation for medical question answering. Muse Spark achieved over a 2% higher score than its closest competitor, GPT 5.4.
This performance is attributed to a clinical training dataset developed with over 1,000 physicians, alongside enhancements to the model's architecture and post-training workflow. Meta reports Muse Spark achieves comparable capabilities using significantly less compute power than its predecessor, Llama 4 Maverick, making it highly efficient.
Muse Spark also excels in scientific chart analysis, outperforming rivals on the CharXiv Reasoning benchmark for technical graphs. This visual reasoning capability extends to practical applications, such as estimating calorie counts from grocery shelf images within the Meta AI app.
Further testing across various benchmarks, including code generation and robot navigation, showed Muse Spark performing competitively, often exceeding at least one rival model.
The "Contemplating mode" feature, which utilizes parallel AI agents to break down tasks, further boosts Muse Spark's output quality, reportedly increasing scores by approximately 8% on difficult benchmarks like HLE.
Muse Spark is the first in Meta's planned series of multimodal reasoning models, signaling a path toward increasingly capable AI systems.