Meta releases Muse Spark Open, an 816B-parameter open-weight model
Meta has released the weights of Muse Spark Open, which it calls the most capable US open-weight model to date. That is Meta's own description; the announcement does not say how it was measured or compared. The model is natively multimodal, with 816B total parameters (42B active) and a 1M-token context window. Meta says it is built for coding and agentic work and excels at long-horizon tasks, complex software engineering and multi-step reasoning.
A large part of the announcement is about running the model at home. Meta says Muse Spark Open is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the latest MLX framework. The reasoning: Apple's unified memory lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold the entire model, and the chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice of weights each token needs. In Meta's words, the two machines behave like one supercomputer with a terabyte of unified memory.
Meta says it tested agentic workflows running entirely on-device with Muse Code, powered by Muse Spark Open. One test shipped real iOS changes to Ice Cubes, an open-source Mastodon client: bug fixes, new features and a full build. Another drove DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.
Distribution is broad. API access is available today through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. The model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA silicon. For customization, Meta points to PyTorch's TorchTitan; on Apple silicon, MLX pulls the weights directly from Hugging Face. The open-weight version also ships with custom, auto-tunable Helion GPU kernels in SGLang, which Meta says deliver 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350.
On safety, Meta says open weights carry their own considerations, so Muse Spark Open received additional safety post-training. It says the results show strong safeguards across the risk categories measured. The model was evaluated under the Meta Superintelligence Scaling Framework, which covers the most severe risks, and Meta says it met the thresholds the framework requires for release. Detailed results are in the model card; a full evaluation report is promised as available soon.
The weights come under the Muse Model License, which Meta describes as having limited commercial terms that allow many developers to use the model for free. The announcement has a Key Benchmarks heading, but the text carries no benchmark scores.
Key facts
- Muse Spark Open is a natively multimodal open-weight model with 816B total parameters (42B active) and a 1M-token context window, built for coding and agentic work.
- Meta says it is the first model in its class to run on a desktop, on two Apple M5 Ultra Mac Studios (up to 512 GB of unified memory each) using MLX.
- API access is offered by seven platforms including Together AI, Groq and OpenRouter; weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
- Helion GPU kernels in SGLang give 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350.
- The Muse Model License has limited commercial terms; Meta says many developers can use the model for free, and a full evaluation report is promised soon.
Why it matters
Meta is releasing the weights of a very large model aimed at coding and agents, and pitching it as something that can run on a desk. An 816B-parameter model with 42B active parameters and a 1M-token context is a size normally associated with data-centre hosting. Meta's claim is that two Apple M5 Ultra Mac Studios, pooling a terabyte of unified memory between them, can hold the whole model. Meta also calls it the most capable US open-weight model to date and the first in its class to run on a desktop; both are the publisher's own statements.
Who it affects
Developers building coding assistants and agentic tools get a model they can call through seven hosted API providers or download and run themselves. Owners of high-memory Apple hardware are the target for the local setup. Teams that fine-tune can use PyTorch's TorchTitan. Inference-engine and cloud vendors are directly involved: the model is supported across SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, and offered on four download hubs and cloud catalogues.
How to use it
Hosted access is available today through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are downloadable on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. To run locally on Apple silicon, use MLX, which pulls the weights directly from Hugging Face; Meta also links an MLX adapter for running the model on Mac Studio. The hardware Meta describes is two Apple M5 Ultra Mac Studios. To customize the model, Meta suggests PyTorch's TorchTitan. The weights are under the Muse Model License, with limited commercial terms that Meta says allow many developers to use the model for free; the exact terms are not spelled out in the announcement.
How solid is it
This is a first-party announcement, so every capability claim is Meta's own. The Key Benchmarks section carries no benchmark numbers in the text, and the announcement does not explain how "most capable US open-weight model" or "first in its class" was determined. The Mac Studio demonstrations (shipping iOS changes to Ice Cubes, driving DaVinci Resolve) are described but come without speed figures. The Helion kernel speedups are specific: 1.03-3.95× at kernel level, but only 3% end to end on AMD MI350. The announcement gives no outside reactions or independent verification.
Risks and caveats
Meta says open weights carry their own considerations and that it added safety post-training, with strong safeguards across the risk categories measured. It says the model met the release thresholds of its Meta Superintelligence Scaling Framework. The full evaluation report is still described as coming soon, so those safety claims cannot yet be checked in full. The license is not unrestricted: it carries limited commercial terms, and the announcement does not say what they are. The desktop setup needs two high-end machines, and the announcement does not state the price of the hardware or of API access, nor the quantization or precision used to fit the model.
“the most capable US open-weight model to date”
— Meta, Muse Spark Open announcement