In the early hours of October 16, Meta released the open-weight model Muse Spark Open, with a total of 816 billion parameters, approximately 42 billion parameters activated per Token, support for a 1 million Token context window, and the ability to accept text, image, video, and audio input. The model adopts a sparse mixture-of-experts architecture, is natively multimodal, focuses on coding and Agent tasks, and targets long-horizon execution, complex software engineering, and multi-step reasoning. Meta calls it the most capable American open-weight model to date.
The focus of this release is local deployment. With the latest version of the MLX framework, Muse Spark Open has been optimized for the Mac Studio with the Apple M5 Ultra chip; a single machine's integrated GPU can use up to 512GB of memory, and two machines combined offer roughly 1TB of unified memory, enough to hold the full model, while 1.2TB/s of memory bandwidth is sufficient to support the parameter reads needed per Token.
Meta tested fully locally running Agent workflows with its Muse Code, including completing iOS code modifications and a full package build for the open-source Mastodon client Ice Cubes, as well as operating DaVinci Resolve to review footage frame by frame and crop it for vertical video.
Evaluation results are mixed. On the Artificial Analysis open-weight model intelligence index leaderboard, Muse Spark Open (xhigh) scored 39, below MiMo-V2.6-Pro's 46 and GLM-5.3's 45, and on par with DeepSeek V4.1Flash; it scored 76.4 on Terminal-Bench2.1 and 60 on DeepSWE1.1, respectively below GLM5.3's 67 and Kimi K3's 68, and its 1401 on GDPval-AA v2 also trailed domestic models such as Qwen3.8-2.4T-A95B.
Developers can access the API through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, and download the weights from platforms like Hugging Face. The model supports mainstream frameworks including SGLang, vLLM, TensorRT-LLM, MLX, llama.cpp, and Unsloth,
Optimized for four chip vendors, its custom Helion kernels achieve 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD GPUs. Meta said that thanks to collaboration with Apple's new CEO John Ternus and the Apple team, the model also underwent additional safety post-training, meeting the release threshold of the superintelligence scaling framework.