Meta is not quietly committed to open models. It is loudly, expensively, almost aggressively committed. The company has now released Muse Spark Open, an 800-billion parameter open-weight model optimized specifically for Apple Silicon. The catch? You need 1TB of unified memory to run it locally.
Meta is calling Muse Spark Open “the most capable US open-weight model to date.” That framing is deliberate. The company has been vocal about its view that American labs need to stay ahead of Chinese competitors on open-source AI. Benchmarks released alongside the model show it outperforming GLM 5.3 and Kimi K3, two prominent Chinese models, on agentic tasks and visual reasoning. Whether those benchmarks hold up in real-world use is a separate question, but the competitive signal is clear.
The hardware story here is genuinely interesting. Muse Spark Open is the first model built to run on Apple’s M5 Ultra-equipped Mac Studio. Meta’s explanation of the setup is worth reading: Apple’s unified memory architecture lets each M5 Ultra Mac Studio use up to 512GB of memory, so two machines chained together hold the entire model. Two Mac Studios, in effect, behave as a single system with a terabyte of unified memory. It’s an unusual deployment architecture, and it puts Apple silicon in a different kind of conversation than the company typically courts.
But the price is not casual. Running Muse Spark Open on four 256GB M5 Ultra Mac Studios costs around $38,000. Apple hasn’t even confirmed pricing or a ship date for the 512GB Mac Studio configuration yet, only that it’s coming in late October. So the full cost picture is still incomplete, and it’s already steep.
For individual developers or small teams, this is basically out of reach. But for larger organizations currently spending heavily on API token costs, the math could work. A one-time hardware investment that enables fully local inference, with no per-token billing and no data leaving your infrastructure, has real appeal at the enterprise level. That’s the audience Meta is actually talking to here, even if the open-weight framing sounds democratic.
This also fits a broader pattern. Models are getting larger, and the race to claim “most capable open-weight” status is accelerating. Mistral, Qwen, DeepSeek, and now Meta are all pushing in this direction. What makes Muse Spark Open notable is the Apple-specific optimization, which is a real technical collaboration rather than incidental compatibility. Muse Spark Open is available to download now from Hugging Face.
