Meta Superintelligence Labs has released Muse Spark Open (Max), an open-weight model featuring 816 billion total parameters. The model utilizes a Mixture-of-Experts architecture, activating 42 billion parameters per token. It supports a context window of one million tokens and is natively multimodal, accepting text, images, video, and audio inputs.
The model achieved 1511 points on the Code Arena: WebDev leaderboard, ranking as the top US-based open model in that category. Additionally, it recorded a score of 89.2 on the MCP-Atlas benchmark for tool usage. Meta notes that the model underwent extra safety post-training compared to previous versions.
Designed for complex software engineering and agentic workflows, the model can operate on two paired Apple M5 Ultra Mac Studios via the MLX framework. Cloud access is provided through Baseten, Fireworks AI, Google Cloud Model Garden, and Amazon SageMaker.
The release follows the introduction of the closed-source Muse Spark in April 2026 and Muse Spark 1.1 in July 2026. While Meta positions the new release as its most capable open-weight model, internal benchmarks show mixed performance when compared to international competitors.
