Muse Spark Open is the multimodal model whose full weights Meta released on October 15, 2026, as the open edition of Muse Spark. It has 816 billion total parameters, activates about 42 billion per token, and offers a 1-million-token context window. It reads text, images, video and audio, and answers in text. In its official blog post, Meta positions it for coding and agentic work: long-horizon tasks, multi-step reasoning and fixing real software bugs are the three scenarios the launch keeps returning to. The weights are free to download from Hugging Face starting the same day, with hosted inference through partners such as Together AI, Fireworks, Baseten, Modal and OpenRouter.
Two Mac Studios acting as one desktop supercomputer
The detail travelling furthest from this launch is not the parameter count but how the model runs. Meta worked with Apple to tune it for the M5 Ultra Mac Studio using Apple's latest MLX framework: one machine offers up to 512GB of unified memory, so two of them provide roughly 1TB, just enough to hold the whole model, and the chip's 1.2TB/s memory bandwidth keeps things moving even though only a small slice of parameters is read per token. Meta's demos included Muse Code shipping real iOS changes to an open-source Mastodon client entirely on-device, and driving DaVinci Resolve frame by frame to pick the strongest moments. The 512GB Mac Studio is not on sale yet, and even the 256GB version costs well over ten thousand dollars, so for now the desktop route reads as a direction for studios and teams rather than a weekend experiment. We previously covered Muse Spark co-authoring papers with mathematicians; this release hands the same model family to outside developers.
The licence says open, but read the commercial limits first
The weights ship under the Muse Model License with limited commercial terms. Many developers can use the model for free, but that is not the same as an unconditional right to resell or host it. Teams planning to build it into a paid product or offer it as an API service should check the licence's limits on commercial scale and redistribution line by line, and should not treat open weights as equivalent to an open-source licence. Size is the other practical barrier: the full BF16 weights are about 1.6TB, the compressed NVFP4 version about 530GB, and community quantisations go smaller still, down to machines with 256GB of memory, with accuracy losses each team must measure itself.
Who should try it now, and who can wait
Teams with local data-compliance requirements that need a long-horizon coding agent should run a round of real-task evaluation now, through hosted channels or a quantised build, judging multi-step completion rates rather than single-turn chat scores. Individual developers are better served by wiring it into their workflow through a hosted endpoint first, confirming it actually beats their current model on their own codebase, and only then pricing hardware. Two top-spec desktops cost as much as a long stretch of cloud usage; what local deployment buys back is not just saved API fees but the certainty that data never leaves the building, and that trade-off has to be calculated per business.
