Meta has released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date, and it worked with Apple to make the model practical to run on a desktop rather than a data centre. The release landed on October 15, 2026.
The specification
Muse Spark Open is natively multimodal with 816 billion total parameters and 42 billion active parameters, using a mixture-of-experts architecture, and it supports a 1-million-token context window. Meta positions it for coding and agentic work — long-horizon tasks, complex software engineering and multi-step reasoning.
The notable engineering claim is that it is the first model in its class to run on a desktop setup: two Apple M5 Ultra Mac Studios with 512GB of unified memory each, using Apple’s MLX framework. Meta explains the arithmetic in its own terms — Apple’s unified memory architecture lets the M5 Ultra’s integrated GPU address up to 512GB per machine, so two machines hold the whole model, and the chip’s 1.2TB/s of memory bandwidth is fast enough to feed each token.
What it costs to actually do this
Running this locally is not a budget project. A Mac Studio with the required M5 Ultra chip and 256GB of memory starts at about $10,799, and the 512GB configuration needed to hold the full model was not yet on sale at the time of the announcement — Apple said it was coming later in October. Meta’s framing is that local inference adds a layer of privacy and can save on cloud costs over time; the entry price is the counterargument.
Where to get it
Weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. The model is also available through hosted APIs at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta says Muse Spark Open is free to download under a licence that permits limited commercial use, and that it ships with day-one support in MLX and llama.cpp.
For teams without the hardware, the hosted route is the realistic option; for anyone who already owns high-memory Apple silicon, the local route is now technically supported rather than an experiment.
Why Meta is doing this
Meta’s business remains almost entirely dependent on advertising, and its AI strategy has run on two tracks: proprietary frontier models and open-weight releases. The company released the smaller Muse Glimmer in August 2026 and the proprietary Muse Wonder earlier in the same week as Muse Spark Open. Meta AI chief Alexandr Wang has argued that many powerful open-weight models are too large to run on most personal computers, and that Apple’s high-memory desktops change that calculation.
Meta CEO Mark Zuckerberg publicly thanked Apple CEO John Ternus for the collaboration, which is notable given how often the two companies have clashed. Apple, for its part, published technical documentation on MLX’s ability to split open-source models across multiple machines.
What we did not do
This article did not run the model. Every specification and price here comes from Meta’s own research blog, Apple-related coverage and the platform listings cited below, not from hands-on testing.
Sources
- Meta AI Research, “Introducing Muse Spark Open”: https://research.meta.ai/blog/muse-spark-open
- TechCrunch, “Meta launches an open-weight AI model that runs on Macs”: https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/
- CNBC, “Apple and Meta team up to make Muse AI model run on Mac computers”: https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html
- MacRumors, “Meta and Apple Bring Muse Spark Open AI Model to Mac Studio”: https://www.macrumors.com/2026/10/15/meta-muse-spark-open-mac-studio/
