Meta Platforms Inc. has released the open weights for its Muse Spark AI model, previously proprietary since its April debut. The model features 816 billion parameters, with 42 billion active per token, and supports a context window of up to 1 million tokens. It is designed for coding tasks and long-running agent operations.
The model underwent safety post-training based on Muse Spark 1.2, released in August. Meta tested the system on two Apple Mac Studio desktops equipped with M5 Ultra chips and 1 terabyte of combined memory. The model demonstrated capabilities in multi-file refactoring, debugging, and frame-by-frame video processing.
In performance benchmarks, Muse Spark Open scored 61.8 on JobBench, surpassing Inkling and Nvidia’s Nemotron 3 Ultra. While it trailed Chinese models like Kimi K3 and Qwen3.8 in coding tests, it achieved a score of 89.2 on the MCP-Atlas benchmark for tool usage, outperforming all listed competitors.
The new Muse Model License allows free commercial use for developers, provided they include a 'Built with Muse Spark Open' label. Companies with over $20 million in gross revenue over 12 months must sign a separate agreement to offer the model as a hosted service. Military and critical infrastructure applications are prohibited.
The model is available via Hugging Face and major cloud providers, including AWS, Google, and Microsoft. Seven inference providers, such as Groq Inc. and Fireworks AI Inc., offer hosted access at launch. Meta continues to keep its most advanced systems closed, including the Muse Wonder models.
