Meta released the weights for Muse Spark Open on October 15, describing it as the most capable US open-weight model to date. The natively multimodal model has 816B total parameters, including 42B active parameters, and a 1M-token context window. Meta says it is built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning. The model is available through API platforms and for download from model repositories.

Meta says Muse Spark Open can run on a desktop with two Apple M5 Ultra Mac Studios using the MLX framework. Each Mac Studio can provide up to 512 GB of unified memory for its integrated GPU, while the M5 Ultra has 1.2 TB/s of memory bandwidth. Meta says the two machines can hold the model together. In on-device tests with Muse Code, the model shipped bug fixes, new features, and a full build for Ice Cubes, an open-source Mastodon client. It also drove DaVinci Resolve end to end, reviewing footage frame by frame to identify moments and crop them vertically.

The weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models; hosted API access is offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model runs with SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA chips. Meta says custom, auto-tunable Helion GPU kernels in SGLang deliver 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. Developers can further customize it with PyTorch.

Meta says Muse Spark Open received additional safety post-training because open weights carry safety considerations. The model met the release thresholds in Meta’s Superintelligence Scaling Framework, which evaluates severe risks; detailed results are in the model card, with a full evaluation report to come. Meta is releasing the weights under the Muse Model License, whose limited commercial terms allow many developers to use the model for free. Developers can further tune it with PyTorch’s TorchTitan or run it on Apple silicon through MLX, which downloads the weights from Hugging Face.

Meta released the weights for Muse Spark Open on October 15, describing it as the most capable US open-weight model to date. The natively multimodal model has 816B total parameters, including 42B active parameters, and a 1M-token context window. Meta says it is built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning. The model is available through API platforms and for download from model repositories.

Meta says Muse Spark Open can run on a desktop with two Apple M5 Ultra Mac Studios using the MLX framework. Each Mac Studio can provide up to 512 GB of unified memory for its integrated GPU, while the M5 Ultra has 1.2 TB/s of memory bandwidth. Meta says the two machines can hold the model together. In on-device tests with Muse Code, the model shipped bug fixes, new features, and a full build for Ice Cubes, an open-source Mastodon client. It also drove DaVinci Resolve end to end, reviewing footage frame by frame to identify moments and crop them vertically.

The weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models; hosted API access is offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model runs with SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA chips. Meta says custom, auto-tunable Helion GPU kernels in SGLang deliver 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. Developers can further customize it with PyTorch.

Meta says Muse Spark Open received additional safety post-training because open weights carry safety considerations. The model met the release thresholds in Meta’s Superintelligence Scaling Framework, which evaluates severe risks; detailed results are in the model card, with a full evaluation report to come. Meta is releasing the weights under the Muse Model License, whose limited commercial terms allow many developers to use the model for free. Developers can further tune it with PyTorch’s TorchTitan or run it on Apple silicon through MLX, which downloads the weights from Hugging Face.