Meta launches an open-weight AI model that runs on Macs

Meta has released Muse Spark Open, an open-weight AI model the company describes as its most capable to date in the United States, with an unusual twist: it was designed to run locally on Macs. Meta published evaluation results for the model on Thursday.
The headline specification is size. Muse Spark Open carries 816 billion parameters and supports a context window of 1 million tokens. Parameters are the learned weights that determine a model's behavior, and the parameter count is a rough proxy for capability as well as for how much hardware you need to serve it. A 1 million token context window is a working-memory figure: it is the amount of text, code or other input the model can consider at once, which matters for tasks such as reviewing a large codebase or a long document in a single pass.
The model is natively multimodal, meaning it handles more than one kind of input without bolting on a separate encoder. Meta says it is built for complex coding tasks and agentic labor, the category of work where a model takes multi-step actions on a user's behalf rather than only answering a question. In a published example, Meta paired the model with Muse Code, its AI coding terminal, and pointed it at a Mastodon client, where it produced bug fixes, new features and a full build. Meta also says it deployed the model to operate DaVinci Resolve, the high-end video editing software, and that it was able to review and crop footage.
Running it locally is the constraint that will define who can actually use it. The local requirement is at least 1 terabyte of RAM, and Meta indicates the model runs best on a Mac Studio configured with the M5 Ultra chip. That is a very large memory footprint, and it puts local inference well outside the reach of a typical laptop. Unified memory on Apple silicon lets the GPU address the same pool as the CPU, which is why a high-memory Mac Studio is a plausible host at all, but 1TB is a configuration most developers will not own. The practical split is likely to be two-tier: a small number of teams running the weights on their own high-memory hardware, and everyone else reaching the model over an API.
That second path is open now. Meta says Muse Spark Open is available through a number of different API platforms, though the announcement does not name them or list prices, rate limits or licensing terms. Open weights give users the option to self-host, fine-tune, inspect and audit a model, and they can be more cost-efficient at scale than paying per token; they also remove the risk that a vendor changes or retires a model out from under a product. Against that, a model this large is expensive to serve yourself, and open weights do not by themselves guarantee that evaluation claims hold up on your particular workload. Meta's performance claims here rest on its own published evaluations, so independent benchmarks and real-world reports are the thing to watch.
The local-first framing is notable for Apple silicon. Apple did not respond in the supplied material, but the excerpt states Meta and Apple teamed up for the launch. Historically, the largest capable models have lived in data centers, and local models have been the smaller, distilled ones. If a model at this scale genuinely runs on a single desktop, that changes the calculus for privacy-sensitive work, offline use and latency, since nothing has to leave the machine.
Why it matters: developers who need frontier-class coding and agentic behavior with data that cannot leave their own hardware now have a named option to evaluate, provided they have the memory. For most product teams the near-term impact is through APIs, where the arrival of another large open-weight model increases pressure on pricing and gives teams a fallback if a hosted model changes. The open question, and it is an inference on my part, is whether the 1TB requirement keeps local use a niche demo or whether it signals the direction Apple silicon workstations are heading.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Meta and Apple have teamed up for the launch of the social media giant's newest model, Muse Spark Open, which will run locally on Apple silicon.