Meta Releases Muse Spark Open: What It Is, Who Can Use It

On October 15, 2026, Meta released the weights of Muse Spark Open, an 816-billion-parameter AI model built for coding and agent work. It can be downloaded or called through an API under a license with limited commercial terms. It matters to businesses and developers looking for an alternative to closed models.

What we know

What it is: per Meta's announcement, it is the open-weight version of Muse Spark. It has 816 billion parameters in total, 42 billion of them active at each step, and a 1-million-token context window.
What it's for: Meta pitches it at coding, long-running agent tasks and multi-step reasoning. Meta calls it natively multimodal; Hugging Face lists it as image-and-text to text.
Meta's claim: the company calls it the "most capable" US open-weight model to date. That is Meta's own evaluation; the page points to the model card and says a full evaluation report is coming soon.
Where to get it: weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. API access comes from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
On a desktop: Meta says it is the first model in its class to run on a desktop setup, using two Mac Studios with M5 Ultra chips and Apple's MLX framework, each with up to 512 GB of unified memory.
License: it ships under the Muse Model License, with "limited commercial terms" that Meta says let many developers use it for free. The page doesn't spell out the limits.
Safety: Meta says it evaluated the model under its internal framework and that it met the thresholds required for release.

What changes and what doesn't

What changes is the menu: there is one more large model that can run outside the big clouds, with data kept in-house.

What doesn't change is the scale. It won't run on web hosting or a normal computer: by Meta's own figures you need two high-end machines with 1 TB of memory between them. And open weights don't mean no strings: the license limits some commercial use.

How to tell if it affects you

Must your data stay inside the company? That is the main reason to self-host an open model. If not, an API is usually cheaper than buying hardware.
Will it power a product you sell? Read the Muse Model License first. The "limited commercial terms" are Meta's wording.
Does it work in your language and on your task? The page says nothing about languages. Run 20 to 30 of your real cases through an API before switching.
If your website runs a chatbot or agent, the model is reachable through the same API providers many tools already use, such as OpenRouter. Compare price and quality with what you use today.
Don't buy the headline alone: "most capable" is Meta's measurement. Wait for the full report or test it on your own work.

Sources

Updates: when Meta publishes the full evaluation report or clarifies the license limits, it will be added here with a link.