Summary
Meta released the weights for Muse Spark Open, a multimodal model with 816 billion total parameters and a 1-million-token context window. Its described desktop setup uses two Apple M5 Ultra Mac Studios working together.
Meta released the weights for Muse Spark Open on October 15, 2026. The company describes it as a natively multimodal model with 816 billion total parameters, 42 billion active parameters and a 1-million-token context window, built for coding and multi-step agentic work.
The model is available through hosted inference platforms and for download. Meta’s desktop demonstration is more specific: Muse Spark Open was optimized to run locally across two Apple M5 Ultra Mac Studios using Apple’s MLX framework.
A two-Mac setup for local agentic work
Meta says each M5 Ultra Mac Studio can provide up to 512 GB of unified memory to its integrated GPU. In the paired setup, the two machines can hold the model, which Meta describes as behaving like a system with a terabyte of unified memory. Its announcement gives a memory bandwidth of 1.2 TB/s for the chip. The desktop configuration described is therefore a pair of Mac Studios working together, rather than a single Mac running the full model.
The company reports testing agentic workflows entirely on-device with Muse Code, its coding tool powered by Muse Spark Open. In one example, it used the model to make and build iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes and new features. In another, it used DaVinci Resolve to review footage frame by frame, identify strong moments and crop them vertically. These are company-reported demonstrations of tasks completed with the model and software, not a general performance guarantee for every workflow.
Access, compatibility and safety
Muse Spark Open is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta lists weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. It says the model works with inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA hardware.
Meta also reports custom, auto-tunable Helion GPU kernels in SGLang. It gives kernel-level speedups of 1.03–3.95× on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. The distinction matters: the larger range refers to individual kernels, while the overall speedup reported for the MI350 is 3%.
The weights are released under Meta’s Muse Model License, which the company says has limited commercial terms while allowing many developers to use the model for free. Meta says it added safety post-training for the open-weight release and that the model met the release thresholds in its Meta Superintelligence Scaling Framework. The company points to the model card for results and says a full evaluation report is forthcoming.