Meta just released the weights for its most powerful AI model yet, and you don't need a data center to run it. Muse Spark Open, an 816-billion-parameter model built for coding and agentic work, can run locally on two Mac Studio desktops — meaning the model stays on your machine, not some cloud server a Big Tech gatekeeper controls.

This is the arrangement Silicon Valley's closed-model oligarchs fear: capable AI that independent developers can download, modify, and run without paying rent to OpenAI or Anthropic for API access. Meta calls it "the most capable US open-weight model to date." The weights are available now on Hugging Face and through Amazon, Google, and Microsoft's model catalogs.

The breakthrough here is local inference at scale. Muse Spark Open uses a mixture-of-experts architecture where only 42 billion of its 816 billion parameters activate per token. Apple updated its MLX machine-learning framework to optimize the model for Mac Studios with M5 Ultra chips. Two Mac Studios — each packing 512 GB of unified memory — link over Thunderbolt 5 to behave as one machine with a terabyte of memory. Meta's AI chief Alexandr Wang said the collaboration brings "leading Agent capabilities to your desktop for the first time."

Meta tested it in real workflows: shipping bug fixes and features to an open-source Mastodon client, and driving DaVinci Resolve end-to-end to scan footage frame by frame and crop vertical video. This isn't a demo. It shipped real code.

The model more than doubled the JobBench agent score of its nearest U.S. rival (61.8 vs. Inkling's 31.1). Chinese open models still lead on pure coding benchmarks — Kimi K3 scored 85 on Terminal-Bench to Muse Spark Open's 76.4. But Muse Spark Open dominated tool use, scoring 89.2 on MCP-Atlas, which tests whether a model can figure out which of 220 tools across 36 live servers to call and in what order. That beats every rival Meta listed, Chinese models included.

The license deserves scrutiny. It's free for developers who adapt the model or ship it in their own software. But any company selling the model as a hosted service that has crossed $20 million in annual gross revenue needs a separate agreement. Military and critical infrastructure uses are banned outright. The weights come with strings.

CNBC framed the story around the Apple-Meta rivalry thawing — noting Tim Cook and Mark Zuckerberg "clashed for years" over Apple's 2021 privacy changes that Meta estimated would cost it $10 billion. With John Ternus now running Apple, the companies found common ground. CNBC also noted Washington's anxiety that "free Chinese models could undercut American AI" — the same politicians who spent years demanding AI regulation now worry that regulation handicaps U.S. firms against open Chinese competitors. Cult of Mac and 9to5Mac mostly explained the technical setup for Apple users. SiliconANGLE provided the benchmark details the others buried.

The real story isn't a corporate détente. It's that the most capable open-weight model from a U.S. lab can now run on hardware a working developer can own — no cloud subscription, no API key, no gatekeeper. The closed-model shops built their business on the assumption that only they could serve capable AI. That assumption just took a hit.

The open question is whether the $20 million revenue cutoff and the military-use ban in Meta's license are modest restrictions — or the thin end of a wedge that lets Meta reassert control once the ecosystem depends on it.