Meta released Muse Spark on Wednesday, a generative AI model natively capable of multimodal understanding; the first model from Meta Superintelligence Labs. According to the official announcement, Muse Spark is can use tools, leverage visual elements as part of its chain of thought and reason through problems by deploying multiple agents in parallel, rather than providing a single agent with more time to work towards a solution. The last feature, known as Muse Spark's "Contemplating" mode, aims to make the model competitive with other frontier models with extreme reasoning modes like Gemini Deep Think and GPT Pro without substantially increasing its response times.

As expected, model performance is mostly reported using benchmark scores. Looking at Meta's findings, Muse Spark does seem to excel at multimodal reasoning: its scores are competitive with those of Opus 4.6, Gemini 3.1 Pro, GPT 5.4, and Grok 4.2 in visual understanding benchmarks including CharXiv Reasoning, MMMU Pro and ScreenSpot Pro. Muse Spark also performs strongly in health-related benchmarks, which is expected, given that Meta seems to be the latest AI company to make health and wellness an area of interest for LLM-powered solutions. According to the company, Muse Spark was trained in collaboration with over 1,000 physicians who curated the training data that enables the model to provide more accurate and detailed responses on areas like nutrition and fitness.

Muse Spark is available now on meta.ai and the Meta AI app, with the "Contemplating" mode coming soon. Meta also plans to make an API preview available to select users.