Meta releases Muse Spark Open, an 816B open-weight model that runs on two Macs Other

Meta released the weights of Muse Spark Open, an 816B-parameter multimodal model built for coding and agents, which Meta calls "the most capable US open-weight model to date". It runs locally on two Mac Studios with M5 Ultra, and is available today on Hugging Face and through API partners.

In Meta's table, Muse Spark Open scores highest of all listed models on JobBench (61.8), MCP-Atlas (89.2), HLE-Diamond without tools (26.1), AA-Omniscience (25), ZeroBench (50.0) and SimpleVQA (73.6). Kimi K3 and GLM 5.3 still score higher on SciCode, Terminal-Bench 2.1, DeepSWE 1.1 and GDPval-AA v2, and Kimi K3 on BabyVision. Meta announced the release on October 15, 2026.

What it is

• Size. 816B total parameters, 42B active per token, and a 1M-token context window.
• Natively multimodal, tuned for coding, agents, long-horizon tasks and multi-step reasoning.
• On your desk. With MLX, two Mac Studios with M5 Ultra hold the whole model: each gives the GPU up to 512 GB of unified memory, so together they have about a terabyte.
• Demos. Running fully on-device in Muse Code, the model shipped bug fixes and features to the open-source iOS app Ice Cubes, and drove DaVinci Resolve end to end, reviewing footage frame by frame and cropping the best moments to vertical.
• Inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. SGLang ships with Meta's Helion GPU kernels for NVIDIA and AMD.

Where to get it

Price

Sources

• Meta: Introducing Muse Spark Open
• Muse Spark Open on Hugging Face