Meta has released an open-weight AI model it describes as the most capable of its kind in the United States, and the company says the model was designed to work with Macs. The model is called Muse Spark Open, and Meta announced evaluation results for it on Thursday, according to TechCrunch.
The headline specification is scale: Muse Spark Open uses 816 billion parameters and a 1 million token context window, per the report. Meta claims the model outperforms other popular open-weight models based on the evaluation results it released. Those results come from Meta itself, not from an independent evaluation, and the supplied evidence does not include the benchmarks, the comparison set or the methodology behind the claim.
Meta states that, beyond its raw size, the model is natively multimodal. Complex coding tasks and agentic labor are what it was constructed to handle. Running locally on Apple computers is another capability it was given, a detail unusual for a model of this size and the one most likely to matter to developers who prefer keeping work on their own hardware.
Local operation comes with a steep hardware floor. Running the model on a Mac requires at least 1 terabyte of RAM, according to Engadget as cited in the TechCrunch report. Meta says the model runs best on a Mac Studio equipped with the M5 Ultra chip. That combination puts local use well outside the reach of a typical freelance laptop setup, and the evidence does not include any smaller quantized variant, a hosted fallback for local users, or pricing for the hardware configuration.
Meta paired the model with Muse Code, its AI coding terminal, and says it deployed the pair across several agentic workflows. One cited deployment involved coding fixes for a Mastodon client, covering bug fixes, new features and a full build. Another involved running DaVinci Resolve, the high-end video editing application, where Meta says the model reviewed and cropped footage.
Those examples are vendor-reported deployments rather than independently reproduced tests. The evidence does not describe how long the tasks took, how much human correction was needed, what hardware was used for each deployment, or whether the results were reviewed by anyone outside Meta. Freelancers evaluating the claims should treat them as Meta's own account of what the model can do.
The report also restates the general case for open-weight models: they can be significantly more cost-efficient than closed and proprietary alternatives, and they can offer transparency and flexibility that closed source models do not allow. That framing is presented as an industry-level advantage rather than a measured result for this specific model, and no cost figures or licensing terms for Muse Spark Open appear in the supplied evidence.
The most concrete practical detail is availability. According to the report, the new model is being made available by Meta right now via a number of different API platforms. Those platforms are not named in the evidence, nor does it state whether the weights themselves can be downloaded; a license, a price per token, and a rate limit are also absent.
For working developers, the API route is the realistic near-term path. A model with an 816 billion parameter footprint and a 1TB RAM local requirement is not something most freelancers will run on a client project machine, so the practical question is what the hosted endpoints cost and how they handle data. Neither answer is in the evidence.
For designers and motion freelancers, the DaVinci Resolve example is the most relevant signal, because footage review and cropping sit close to real editing work. But a single vendor-reported deployment is not a workflow endorsement, and the evidence does not say whether the model touched timeline editing, color, rendering or anything beyond reviewing and cropping clips.
The coding claims deserve similar caution. A Mastodon client fix that included bug fixes, new features and a full build is a meaningful demonstration if accurate, but the evidence gives no repository, no diff, no test results and no indication of how much of the work was model-driven versus human-directed. Agentic coding claims are exactly the kind that need reproducible detail before they change anyone's tooling decisions.
The Mac angle is the most distinctive part of the announcement. Designing a model of this scale to run locally on Apple hardware is a notable positioning choice, and it points at users who want inference on their own machine. The 1TB RAM requirement, however, means that positioning currently applies to a narrow slice of buyers rather than to the freelance mainstream.
What remains unknown is substantial. There is no independent benchmark, no named API partners, no pricing, no license terms, no information on context-window performance at the stated 1 million tokens, and no detail on multimodal inputs beyond the word natively. The evaluation results Meta released are the only performance evidence in the supplied material.
The reasonable read for this audience is to treat Muse Spark Open as a model to watch through its API rather than one to plan local infrastructure around. If the hosted endpoints are priced competitively and the coding performance holds up outside Meta's own testing, it becomes a candidate for agentic coding and media tasks. Until pricing, licensing and independent evaluation appear, the local Mac story is a specification, not a workflow.
Meta's framing of the model as the most capable open-weight model in the United States is also a scoped claim. It is limited by geography in Meta's own wording, and the comparison set behind it is not described in the evidence. Readers should not read it as a global ranking or as a settled result.