Meta's New Open-Weight AI Model Is Built to Run on Your Mac

Meta released a new open-weight AI model this week that it's calling the most capable one available in the United States, and it comes with an unusual selling point: it's designed to run locally on Apple hardware.

The model, called Muse Spark Open, carries 816 billion parameters and a 1 million token context window, according to evaluation results Meta published on Thursday. In plain terms, that context window means the model can hold roughly a million tokens worth of text, code, or other input in its working memory at once, which matters a lot for tasks like digesting an entire codebase or a long video project in a single pass. Meta says the model outperforms other popular open-weight models on its benchmarks, though as with any vendor-run evaluation, that claim comes from Meta's own testing.

Open-weight models differ from closed systems like the proprietary models that power many commercial chatbots because their underlying parameters, essentially the learned settings that determine how the model behaves, are published for anyone to download and run. That openness tends to make them cheaper to deploy, easier to inspect, and more flexible to customize than closed alternatives, which is part of why companies like Meta keep investing in them even as rivals lean into closed, subscription-based products.

What sets Muse Spark Open apart isn't just the specs. It's natively multimodal, meaning it can work across different types of input like text and images without bolted-on workarounds, and Meta built it specifically with coding and agentic work in mind. Agentic, in this context, refers to AI systems that don't just answer questions but actually carry out multi-step tasks on their own, like writing code, testing it, and fixing what breaks.

Running it won't be easy

The catch is hardware. Running Muse Spark Open locally requires at least 1 terabyte of RAM, according to Engadget's reporting, which rules out virtually every consumer Mac on the market today. Meta says the model performs best on a Mac Studio equipped with Apple's M5 Ultra chip, Apple's top-tier desktop processor built for exactly this kind of heavy computational load.

That's a steep barrier for casual users, but it's clearly not who this release is aimed at. For studios, development shops, and technical teams willing to invest in high-end Mac Studio hardware, the appeal is being able to run a frontier-class model entirely on-device rather than routing requests through the cloud, which can matter for cost, latency, and keeping sensitive data in-house.

Coding terminals and video edits

Meta paired the new model with Muse Code, its own AI coding terminal, and ran it through a handful of real-world test cases to show what it can do. In one example, the company had it work on a Mastodon client, handling what Meta described as bug fixes, new features, and a full build cycle, essentially letting the model operate as something close to an autonomous engineer rather than just a code suggestion tool.

Meta also pointed to a less expected use case: video editing. The company deployed Muse Spark Open inside DaVinci Resolve, the professional-grade editing software widely used in film and broadcast work, where the model was able to review footage and handle cropping tasks. That's a meaningful signal that Meta sees this model's usefulness extending well beyond text generation and into creative production workflows that have traditionally required a human editor's eye.

Muse Spark Open is available now through a range of API platforms, Meta says, giving developers and businesses multiple paths to access it even if they don't have the local hardware to run it themselves.

The launch fits into a broader pattern playing out across the AI industry, where open-weight releases have become a competitive front of their own. Companies gain goodwill, developer mindshare, and real-world testing data by putting powerful models into the hands of anyone who can run them, even when the practical hardware requirements keep that pool relatively small for now. As chips like Apple's M5 Ultra become more common in high-end desktop setups, the gap between what's technically possible and what's actually accessible on a Mac sitting on someone's desk is likely to keep shrinking.