Meta introduces Muse Spark, the first model from its new advanced AI lab

Meta has introduced Muse Spark, the first model in the Muse family developed by Meta Superintelligence Labs (MSL), a division created by the company to consolidate its advanced artificial intelligence efforts. The model combines in a single system the ability to understand text and images, execute external tools, reason visually step by step, and coordinate multiple AI agents working in parallel.

Meta positions Muse Spark within a long-term scaling strategy. The company presents it as the first model in a series that will grow in capability, announcing investments across the entire development chain, from research and model training to infrastructure, including the new Hyperion data center.

According to benchmark results published by Meta, the model shows an uneven profile across categories. It performs particularly well in multimodal tasks, leading in tests such as CharXiv Reasoning (86.4) and ERQA embodied reasoning (64.7), and in health, scoring 42.8 in HealthBench Hard compared to 14.8 for Claude Opus 4.6 and 20.6 for Gemini 3.1 Pro. In abstract reasoning, however, there is more room for improvement, with a 42.5 in ARC AGI 2 versus 76.5 for Gemini 3.1 Pro. Results in agentic tasks are competitive but without clear leadership.

For more complex tasks, Meta has developed Contemplating mode, which runs multiple agents reasoning simultaneously in parallel. In this mode, the model reaches 58% on Humanity's Last Exam and 38% on FrontierScience Research, figures that place it in line with the most advanced reasoning modes of models such as Gemini Deep Think or GPT Pro. This mode will be rolled out gradually on meta.ai.

In the health domain, Meta collaborated with more than 1,000 physicians to prepare specific training data, aiming for the model to provide more accurate responses on health-related topics, such as the nutritional content of foods or the muscles involved in different physical exercises.

From a technical standpoint, Meta states that Muse Spark requires significantly less computing power than its predecessor Llama 4 Maverick to reach an equivalent level of performance. The company attributes this improvement to changes in the model architecture, optimization processes, and training data curation.

On safety, Meta commissioned external evaluations from Apollo Research, which found that Muse Spark shows a high ability to identify when it is being evaluated. The company states that, while this behavior warrants further research, it does not consider it an obstacle to the model's release.

Muse Spark is available today on meta.ai and in the Meta AI app. The company has also opened a private API preview for selected users.

Key points

• Muse Spark is the first model in the Muse family, developed by Meta Superintelligence Labs (MSL).
• It marks the start of a scaling roadmap toward increasingly capable models.
• The model natively processes text and images and can coordinate multiple AI agents.
• It leads in multimodal and health benchmarks, but has room to improve in abstract reasoning and agentic tasks.
• Contemplating mode, with parallel agents, achieves results comparable to the top models on the market.
• Collaboration with more than 1,000 physicians to improve health-related responses.
• Requires far less computing power than its predecessor Llama 4 Maverick for equivalent performance.
• An external evaluation found the model can identify when it is being tested.
• Available today on meta.ai, with private API access for selected users.