Meta released Muse Spark Open on Thursday, October 15, calling it the most capable United States open-weight model to date. That is a vendor claim, and independent testing will settle it, but the release itself is the news: after focusing on its closed Muse Wonder models earlier this week, Meta has put a frontier-scale model into downloadable weights that developers can host, inspect, and adapt outside Meta's own service.
The specifications explain the ambition. Muse Spark Open is a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active per token, a one million token context window, configurable reasoning, and native tool calling. It accepts text, images, documents, audio, and video, and returns text. Meta built it for coding agents and long-horizon work: multi-file refactoring, extended debugging, and multi-step tasks that run long after a single prompt would have been forgotten.
The strangest and most interesting part is where Meta expects it to run. The model is optimized for Apple's Mac Studio with the M5 Ultra chip, using Apple's MLX framework. Two linked Mac Studios, each addressing up to 512 gigabytes of unified memory, can hold the whole model between them. Meta says it tested that setup with Muse Code shipping real changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve to review and crop footage. Local still means serious hardware, not a laptop, but it is a desktop, not a data center.
Availability is broad from day one. The weights are on Hugging Face, AMD announced day-zero support on its Instinct GPUs with vLLM, Amazon added the model to SageMaker JumpStart, and hosted providers are offering API access. Read the license before building a business on it, though: this is the Muse Model License with limited commercial terms, not unrestricted open source, and large API providers need a separate agreement with Meta. Downloadable is a real advantage, but it is not the same as unconditional.
For builders, the practical choice splits in two. Most teams will meet Muse Spark Open through a hosted API, where someone else pays for the memory and the power bill, and will judge it the unromantic way: does it finish coding and agent tasks accurately on the first attempt, and what does each completed task cost? A smaller set of teams, the ones with sensitive code, strict data rules, or steady heavy workloads, now has a credible local option that did not exist yesterday at this scale. Meta also concedes the ceiling: its own comparisons put Spark Open well ahead of rival open models in several categories, but still short of its paid Muse Wonder Full models.
That gap is the strategy. Meta can give away a very large model, grow the ecosystem around its tools and license, and still keep its best system behind the counter. Competitors selling only closed APIs should feel the pressure anyway. When an open model is good enough for real agent work, price and control start beating leaderboard bragging rights, and that is a fight the closed labs will have to have on builders' terms.
