Meta releases Muse Spark Open, an 816B-parameter open-weight AI model

Meta has released the weights for Muse Spark Open, a multimodal open-weight model designed for coding, complex reasoning, and agent-based workflows. Meta announces in an official post on Meta AI Research that the model has 816 billion total parameters, with 42 billion active for each response, and supports a context window of up to one million tokens.

The release targets organizations and developers that want to run a capable model on their own infrastructure rather than relying entirely on a cloud service. Meta says Muse Spark Open can run locally on two Apple Mac Studio systems with M5 Ultra chips, using Apple’s MLX machine learning framework.

That setup depends on Apple’s unified memory architecture. Meta says each M5 Ultra Mac Studio can provide up to 512GB of shared memory, allowing two systems to hold the full model. The company describes the configuration as a desktop-scale system with roughly one terabyte of unified memory.

Availability and tools

Developers can download Muse Spark Open through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted access is also available through providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Meta says the model works with major inference tools such as vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth. It also supports further tuning through PyTorch and TorchTitan.

The company positions Muse Spark Open for long tasks, including software engineering and workflows that require multiple steps. Meta reports that it tested the model on coding work for the open-source Mastodon client Ice Cubes and on video review and cropping tasks in DaVinci Resolve. Those examples are Meta’s own tests and do not independently establish performance across broader production use.

Muse Spark Open uses the Muse Model License, which Meta says permits many developers to use the model free of charge, while imposing limited commercial terms. Meta also says it conducted safety evaluations under its Superintelligence Scaling Framework and added safety-focused post-training before releasing the weights.