Meta has released the weights for Muse Spark Open, and the company says the model runs locally on a pair of M5 Ultra Mac Studios. It is a natively multimodal model with 816B total parameters, 42B of them active, and a 1M-token context window.
In its research post, Meta calls Muse Spark Open the most capable US open-weight model to date. Mark Zuckerberg posted on X that the weights are out today and that the model runs well on the new Mac Studios, crediting work with Apple and its CEO, John Ternus.
Ternus replied on X less than an hour later. He congratulated the team and called it amazing to run a model of this caliber and scale so efficiently from a desk, adding that Mac Studio keeps pushing the boundaries of on-device AI.
@finkd Congrats to the team! Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!
— John Ternus (@johnternus) October 15, 2026
Two Mac Studios hold the whole 816B model in memory
Each Mac Studio with M5 Ultra can give its GPU up to 512GB of unified memory. Two machines add up to 1TB, which is enough to load every parameter. The chip’s 1.2TB/s of memory bandwidth is fast enough to read the 42B parameters each token needs, so the pair behaves like one small supercomputer.
Apple’s MLX framework does the work here, and it pulls the weights straight from Hugging Face. Meta also lists support for llama.cpp, Unsloth, SGLang, vLLM, TensorRT-LLM and Hugging Face Transformers. The memory headroom is the same one behind the 4.3x AI performance jump over the M3 Ultra.
Meta tested agent workflows that ran entirely on the Macs, using Muse Code powered by Muse Spark Open. In one test the model shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build. In another it drove DaVinci Resolve from start to finish, reviewing footage frame by frame to find the strongest moments and crop them vertically.
Muse Spark Open leads on agentic tests but trails on coding
Meta’s benchmark chart compares the model with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B, Mistral Large 4 and others. Muse Spark Open posts the highest scores in the chart on MCP-Atlas (89.2), JobBench (61.8) and HLE-Diamond without tools (26.1). It also leads the multimodal ZeroBench and SimpleVQA tests with tools.
Coding is a weaker area for the model, with 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3 and 83.9 for GLM 5.3. On DeepSWE 1.1 it scores 60, against 68 for Kimi K3. Kimi K3 also leads BabyVision, 85.7 to 80.2. These are Meta’s own numbers, and independent testing has yet to confirm them.
Weights are available now under a limited commercial license
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted API access comes from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
The license is the Muse Model License, which has limited commercial terms and lets many developers use the model for free. Because the weights are open, Meta added extra safety post-training and says the model met the release thresholds in its Meta Superintelligence Scaling Framework. The model card has the details, and a full evaluation report is coming.
The release also ships custom, auto-tunable Helion GPU kernels in SGLang, which Meta says deliver 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD GPUs.
Our Take
Running an 816B model on hardware you own is the strongest argument yet for maxing out the memory on an M5 Ultra Mac Studio. The two-machine requirement keeps this in developer and small-team territory, not something most Mac owners will try.
