Muse Spark Open Released as Most Capable US Open-Weight Model
Developers gain access to an 816-billion-parameter open-weight model with a 1-million-token context window, designed for coding and long-running agentic tasks, and runnable for the first time on desktop hardware using two Apple M5 Ultra Mac Studios. The weights for Muse Spark Open are available now on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, with hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
The release marks the first time a model in this class runs entirely on a desktop, powered by Apple’s MLX framework and the unified memory architecture of the M5 Ultra chip.
What makes Muse Spark Open different
Muse Spark Open is a natively multimodal model with 816 billion total parameters, 42 billion of which are active for any given token, and a 1-million-token context window. That combination targets long-horizon work: multi-step reasoning, complex software engineering, and agentic workflows that span many turns and many tools.
The model was optimized through the latest MLX framework to run on Mac Studio with M5 Ultra. Apple’s unified memory architecture lets the integrated GPU use up to 512 GB per Mac Studio, so two machines together can hold the entire model. The chip’s 1.2 TB/s of memory bandwidth is enough to read the small slice each token needs, and the result is that two Mac Studios behave like one supercomputer with a terabyte of unified memory.
Agentic workflows tested on-device
Testing covered a range of agentic workflows running entirely on-device, using Muse Code powered by Muse Spark Open. The work included shipping real iOS changes to Ice Cubes, an open-source Mastodon client: bug fixes, new features, and full builds. The model also drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically.
Where to get the model and run it
Model weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted inference is available through Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter.
The open-weight version ships with custom, auto-tunable Helion GPU kernels in SGLang, delivering 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, and a 3 percent end-to-end speedup on AMD MI350.
Muse Spark Open runs on leading inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Developers can further customize the model using the PyTorch framework, including TorchTitan for additional tuning. To run it on Apple silicon, MLX pulls the weights directly from Hugging Face.
Safety and licensing
The model underwent additional safety post-training for the open-weight release. Evaluation followed the Meta Superintelligence Scaling Framework, covering the most severe risks, and the model met the thresholds required for release. Detailed results are in the Muse Spark Open model card, with a full evaluation report to follow.
Weights are released under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.
FAQ
What is Muse Spark Open?
Muse Spark Open is an open-weight model with 816 billion total parameters (42 billion active) and a 1-million-token context window. It is natively multimodal and built for coding and agentic work.
How can Muse Spark Open run on a desktop?
It is optimized through Apple’s MLX framework to run on Mac Studio with M5 Ultra. Two M5 Ultra Mac Studios together provide a terabyte of unified memory, enough to hold the full model and run agentic workflows on-device.
Where can developers download and run Muse Spark Open?
Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth.