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

TechCrunch reports Meta launched Muse Spark Open, calling it the most capable open-weight model in the United States. The model has 816 billion parameters and a 1 million token context window, is natively multimodal, and is built for complex coding and agentic work. Meta designed it to run locally on Apple computers; Engadget says local runs need at least 1TB of RAM, with best results on Mac Studio with M5 Ultra.

Meta paired the model with Muse Code, its AI coding terminal, and described agentic demos including Mastodon client fixes and using DaVinci Resolve to review and crop footage. The model is available now through multiple API platforms.

This is Meta doubling down on open weights as a distribution strategy — and pairing it with Apple silicon so “open” doesn’t only mean “someone else’s GPU cloud.”

For teams that can clear the memory bar, a coding-optimized multimodal MoE with a million-token window is a real workstation story: screenshots, voice notes, and long repos in one loop. For everyone else, the API platforms matter more than the Mac demo.

We’re bullish on open weights that ship with serious tooling. We’re also clear-eyed: 1TB RAM gatekeeps the local dream, and agent demos on video editors and social clients will get stress-tested in the wild. I’m watching third-party evals versus Meta’s own numbers, and whether Muse Code becomes the default harness people actually use.

Open-weight models can be more cost-efficient and flexible than closed APIs; Meta is pushing Muse Spark Open as both a local and API-accessible release.