Meta releases Muse Spark Open, an open-weight AI model built to run on two Mac Studios

Meta released Muse Spark Open on 15 October 2026, making the weights of a large new AI model available for developers to download and run on their own hardware. The company describes it as “the most capable US open-weight model to date” and says it is designed for coding and so-called agentic work, in which an AI carries out long, multi-step tasks such as fixing software or editing files with little supervision.

Muse Spark Open is the open-weight version of Muse Spark, Meta's closed flagship model. It has 816 billion parameters in total, but only about 42 billion are active at any one time, which reduces the computing power needed to run it. It can take in 1 million tokens of context, enough for very large codebases or long documents, and Meta says it is natively multimodal.

Running it on a desk

Meta tuned the model for Apple hardware. According to the company, it runs on two Mac Studio computers with Apple's M5 Ultra chip using Apple's MLX framework, with each machine offering up to 512 GB of unified memory. TechCrunch, citing Engadget, noted that running it locally needs at least 1 TB of memory, which puts it far beyond ordinary laptops.

For most developers, access will come through the cloud. Meta says the weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, and that API providers including Groq, Fireworks AI, Together AI, OpenRouter and Baseten serve the model. It is released under the Muse Model License, which Meta says includes limited commercial terms, so businesses will need to check the conditions before building on it.

TechCrunch reported that Meta paired the model with Muse Code, its AI coding terminal, and showed it fixing bugs and adding features in a Mastodon app and reviewing and cropping footage in the DaVinci Resolve video editor.

Safety and caveats

Because published weights cannot be withdrawn, Meta said it gave the model additional safety training and evaluated it under its Meta Superintelligence Scaling Framework, concluding that it met the thresholds for release. It promised a full evaluation report soon. The headline performance claims come from Meta's own testing, so independent benchmarks will be needed to show how the model compares with rival open models.

Why it matters

Open-weight models let companies run AI on their own infrastructure, which can lower costs and keep sensitive data in-house. Meta's release adds a large US-built option to that market at a time when many leading AI labs offer their most powerful systems only as paid online services.