Meta has released the weights of Muse Spark Open, an AI model built for coding and agent-style tasks, which the company calls the most capable open-weight model from the US to date. The model became available on October 15 through Hugging Face, major cloud platforms and several API providers.
What the model offers
According to Meta's research blog, Muse Spark Open has 816 billion parameters in total, with 42 billion active at any one time, and a context window of one million tokens. The company says it is natively multimodal and designed for coding and agentic work, including long-running tasks, complex software engineering and multi-step reasoning. A post on the PyTorch blog adds that the open version received extra safety training so that it is harder to misuse when run on its own, without guardrails built around it.
Running it on Mac Studios
Meta says it worked with Apple to get the model running locally on the M5 Ultra Mac Studio, using two machines together through MLX, Apple's machine-learning framework for its own chips. In a post on X, chief executive Mark Zuckerberg said the model performs well on the new Mac Studios because of that collaboration. Users can link two to four Studios for coding or agentic work.
Where to get it
Developers can reach the model through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, Meta said. The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. The model works with inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth, in addition to MLX. Meta says it has optimised the release for hardware from AMD, Apple, Intel and NVIDIA.
