Meta Frees Muse Spark: 816B Weights Run on Desktops
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
Meta released the weights of Muse Spark Open, an 816B-parameter open-weight model that can run entirely on two Apple M5 Ultra Mac Studios on a desk.
What happened: Meta ships 816B open weights
On October 15, 2026, Meta released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date.
The model is natively multimodal with 816 billion total parameters (42 billion active), a 1M-token context window, and is built for coding and agentic work — Meta says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, with API access via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Why it's surprising: a desktop frontier model
The first-in-class detail is the hardware recipe: 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.
Each Mac Studio's integrated GPU can use up to 512GB of unified memory, so two machines hold the entire model at 1.2 TB/s bandwidth — Meta says they behave like one supercomputer with a terabyte of unified memory.
PYMNTS reports Zuckerberg credited work with Apple CEO John Ternus, and The Wall Street Journal put the Mac setup cost at about $40,000 — steep, but against thousands of dollars a month in cloud rental.
The evidence: demos and benchmarks
Meta demonstrated on-device agentic workflows: Muse Code shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build, and drove DaVinci Resolve end to end, reviewing footage frame by frame to crop vertical moments.
On benchmarks, VentureBeat reports Meta's own table puts Spark Open ahead of listed rivals on JobBench (61.8) and MCP-Atlas (89.2) but behind GLM 5.3 and Kimi K3 on Terminal-Bench 2.1 and DeepSWE 1.1, with a reported Intelligence Index of 39.
OrcaRouter notes Artificial Analysis' hosted listing independently scores it 39 at xhigh — ninth of 118 open-weight models above 150B parameters.
The fine print: license and gated access
The weights ship under the Muse Model License with limited commercial terms, free for many developers but with strings: VentureBeat reports a $20 million provision requiring notification and a written agreement with Meta for model-as-a-service businesses, plus a required 'Built with Muse Spark Open' attribution.
The Hugging Face gate asks for full legal name, date of birth and full organization name with corporate identifiers, and the form cannot be edited after submission.
Meta also says the model received additional safety post-training and was evaluated under its Superintelligence Scaling Framework, with a full evaluation report still pending.
The stakes: the China open-weight gap
The release lands amid a candid verdict on the open-weight race.
Rest of World reports Chinese models took four of the top five spots on OpenRouter's leaderboard last week, and Stanford DigiChina's Graham Webster said that if you need relatively frontier capability, right now that's Chinese open models only.
Hugging Face's March 2026 assessment found Chinese models took 41% of downloads over the prior year.
Whether Muse Spark Open — with strong safeguards and Western provenance — shifts adoption among businesses and governments seeking alternatives to both US API dependence and Chinese dominance is the open question.
What comes next: run it yourself
Builders can start today: weights are downloadable from Hugging Face, MLX pulls them directly for Apple silicon, and day-one support spans SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp and Unsloth across AMD, Apple, Intel and NVIDIA hardware.
The open-weight release ships with custom auto-tunable Helion GPU kernels in SGLang, delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs per Meta. Developers can fine-tune via PyTorch's TorchTitan, and OrcaRouter notes BF16 (about 1.5TB) and NVFP4 builds are available.
With only one independent benchmark so far, real-world developer testing will decide whether the 'most capable US open-weight' claim holds.
Sources
- research.meta.ai › Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date
- theguardian.com › Meta releases Muse Spark Open, a new 'open-weight' AI model
- pymnts.com › Meta and Apple Bring Open-Weight Frontier Model to Desktops
- venturebeat.com › Meta unveils Muse Spark Open, the new top U.S. open weights model
- orcarouter.ai › Muse Spark Open: 816B Weights, and the Licence Fine Print
- restofworld.org › Western open-weight models still can't beat China