Blog
Meta releases Muse Spark Open, an 816B open-weight model that can run on two Mac Studios
On 15 October 2026 Meta released the weights for Muse Spark Open, a multimodal mixture-of-experts model with 816 billion total parameters and a one-million-token context, built for coding and agents. Meta calls it the most capable US open-weight model; that ranking is Meta's own claim.
Meta released Muse Spark Open on 15 October 2026, an open-weight model it describes as "the most capable US open-weight model to date." That ranking is Meta's own claim, based on evaluation results the company published alongside the release. What can be checked independently is what was shipped: the weights are downloadable, and the model is available through several hosting providers from day one.
What Meta released
According to Meta's announcement, Muse Spark Open is a natively multimodal mixture-of-experts model with 816 billion total parameters, of which about 42 billion are active for each token. It has a context window of up to one million tokens. Hugging Face's launch post describes the same architecture and notes that the model takes text, images, audio and video as input and produces text as output.
Meta says the model was trained mainly for coding and agentic work: reading a codebase, writing and debugging code, calling tools and carrying out long, multi-step tasks. Because it accepts several kinds of input, Hugging Face points out that the same model can look at a screenshot, listen to a voice note or watch a screen recording while working on code.
Hugging Face also places the model in Meta's wider Muse family. It follows Muse Glimmer, a smaller dense model Meta released earlier that was designed for running locally. Muse Spark Open is Meta's first release in the family to use a mixture-of-experts design at this scale.
Running it on a Mac
The detail that drew the most attention is local use. Meta says Muse Spark Open is the first model in its class that can run on a desktop setup of two Apple Mac Studio computers with the M5 Ultra chip, using Apple's MLX framework. Hugging Face says the model works in oMLX and mlx-lm, with 3-bit and 4-bit quantised versions available.
This is not a typical home setup. TechCrunch, citing Engadget, reports that running the model locally needs at least 1 terabyte of memory. For most developers, hosted access will be the practical route.
Where it is available
Meta lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta says the model is supported in common inference engines, including vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth.
Meta's demonstrations
Meta showed the model working with Muse Code, its AI coding terminal. In one example the model worked on an open-source Mastodon client, handling bug fixes, new features and a full build. In another it operated DaVinci Resolve, the video editing software, reviewing footage and cropping it. These are company-selected demonstrations, not independent tests.
Why it matters
Open-weight models let organisations run a model on their own hardware, inspect it and fine-tune it, which can matter for cost, privacy and control. Meta's release adds a very large, multimodal, agent-focused option to that category. Its performance claims are still the company's own, and independent benchmark results will give a clearer picture of how it compares with other open and closed models.