Meta released Muse Spark Open on Thursday, October 15, calling it the most capable United States open-weight model to date. Unlike a closed system that stays behind a company API, an open-weight release lets developers download the trained parameters, adapt the model, and run it on their own hardware. Meta is aiming this one squarely at coding and agentic work: long, multi-step software tasks that require reasoning, tool use, and sustained context rather than a single quick answer.
The scale explains both the ambition and the catch. Muse Spark Open has 816 billion total parameters, with 42 billion active at a time, and a context window of up to 1 million tokens. It is natively multimodal, so it can work across different kinds of input as well as text. That combination is built for extended engineering jobs and multi-step reasoning, the kind of work where an agent needs to hold a large codebase, a plan, and a trail of earlier decisions in mind at once.
The most striking deployment story is local, not cloud. Meta worked with Apple to tune the model for Mac Studio desktops powered by the M5 Ultra chip, using Apple's MLX framework. Two machines, each with up to 512 gigabytes of unified memory, can hold the model together and behave, in Meta's telling, like a single desktop supercomputer with a terabyte of memory. Apple has not yet shipped the 512-gigabyte Mac Studio configuration, which is expected later this month, and even the existing high-memory machines are far from inexpensive. Local here means a serious workstation, not a laptop on a coffee table.
Meta demonstrated the setup through Muse Code, its AI coding terminal. In one example, the system shipped changes to Ice Cubes, an open-source Mastodon client, handling bug fixes, new features, and a full build. In another, it drove the DaVinci Resolve video editor, reviewing footage and cropping clips. Those are vendor demonstrations, not independent benchmarks, but they illustrate the intended use: an agent that can operate real software over many steps, on hardware the developer controls, without sending every prompt to a remote service.
Access is broad from day one. Developers can reach the model through API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, and download the weights through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The license allows limited commercial use. For businesses, that menu matters: the same model can be tested through an API, then brought in-house where data control, cost, or customization justify the effort.
There are real trade-offs. Running a model of this size locally demands substantial computing resources and technical expertise, even with only a fraction of its parameters active per token. Meta chief executive Mark Zuckerberg has argued that advanced AI should not be concentrated among a small number of companies, and this release puts that argument into downloadable form. Whether teams adopt it will depend less on the headline parameter count than on how reliably it finishes long coding and agent tasks in the wild, and what it costs to run well outside a demonstration.