A new open-weight model called Muse Spark Open is now available for download and through hosted APIs, bringing agentic coding abilities and a million-token context window to a wider developer audience. The model, released by Meta’s research group, is the most capable US open-weight model released to date and is built natively for multimodal input, software engineering, and long-horizon reasoning tasks.<\/p>\n
Muse Spark Open is a natively multimodal model with 816 billion total parameters, of which 42 billion are active per inference. It carries a one-million-token context window. The model is designed for coding and agentic work, with stated strengths in long-horizon tasks, complex software engineering, and multi-step reasoning.<\/p>\n
Its size and architecture make it a tool for serious software work, not a casual chatbot. An agentic coding session can now plan a multi-file change, hold a large codebase in context, and execute the steps in sequence without losing the thread.<\/p>\n
Muse Spark Open is the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios paired through the MLX framework. Apple’s unified memory architecture lets the M5 Ultra’s integrated GPU use up to 512 GB per machine, so two Mac Studios together can hold an entire 816B-parameter model. The chip’s 1.2 TB\/s of memory bandwidth is fast enough to read the small slice of weights each token needs.<\/p>\n
The practical result is that two Mac Studios behave like one supercomputer with a terabyte of unified memory, which lets a developer run a frontier-class model at home rather than renting cloud capacity.<\/p>\n
On-device testing covered a range of agentic workflows running entirely on two Mac Studios using Muse Code powered by Muse Spark Open. One workflow shipped real iOS changes to Ice Cubes, an open-source Mastodon client, covering bug fixes, new features, and a full build. Another workflow drove DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.<\/p>\n
The model is available now through hosted inference platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Model weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.<\/p>\n
For local or custom deployment, the weights run on leading inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Each engine has optimizations for AMD, Apple, Intel, and NVIDIA silicon. Developers can further customize the model using the PyTorch framework, including TorchTitan for additional tuning.<\/p>\n
The open-weight release 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, plus a 3 percent end-to-end speedup on AMD MI350 hardware.<\/p>\n
Open-weight releases carry their own safety considerations, and Muse Spark Open received extra safety post-training for that reason. Meta evaluated the model under its Meta Superintelligence Scaling Framework, covering the most severe risk categories, and the model met the thresholds required for release. Detailed results are published in the Muse Spark Open model card, with a full evaluation report forthcoming.<\/p>\n
The weights are released under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.<\/p>\n
Muse Spark Open has 816 billion total parameters, with 42 billion active per inference, and a one-million-token context window.<\/p>\n
Yes. It is the first model in its class that runs on a desktop using two Apple M5 Ultra Mac Studios through the MLX framework. Two Mac Studios can hold the full model thanks to up to 512 GB of unified memory per machine.<\/p>\n
Weights are available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.<\/p>\n