816 Billion Parameters on Four Macs: Meta's Strongest U.S. Open Model Comes With a Hardware Bill
An 816-billion-parameter model squeezed into four Mac Studios and generating 36.7 tokens per second is what Meta's AI chief calls the strongest open-weight model in America — and what gives Apple one more reason to sell Macs.
Three key facts
On October 15, Meta released Muse Spark Open. According to 36Kr and the Vietnam News Agency, the model uses a mixture-of-experts architecture with 816 billion total parameters, about 42 billion active per inference, and a 1-million-token context window. Alexandr Wang, who leads Meta AI, called it the most capable open-weight model developed in the United States, though still behind Muse Wonder, the closed model Meta released earlier that week.
Its selling point is running locally on Apple's Mac Studio. Apple used MLX and Thunderbolt 5 to link multiple machines. At 4-bit quantization the model needs about 493GB of memory; two M5 Ultra Mac Studios running tensor parallelism hit about 25 tokens per second, four hit about 36.7, with a first-token wait of about 4.8 seconds. The catch: an M5 Ultra Mac Studio starts at 46,999 yuan, and four of them, plus cables and power, add up to a serious bill.
Meta is betting on open weights in a market Chinese companies already heated up. Meta released the smaller Muse Glimmer only in August; now it is putting out an 816-billion-parameter model with a license that permits commercial use within limits. Qwen, Kimi K3 and GLM are already formidable rivals in the global open-weight market, and Meta needs open weights to pull developers back into its ecosystem.
Wang Dou's Take
Don't mistake "strongest open model in America" for a world championship. It is closer to an MVP award for the American bracket: the trophy is real, but it only counts inside that bracket, while Chinese models have been running laps on the open-weight track. Meta is not releasing 816 billion parameters because it suddenly fell in love with openness. It is doing so because closed models are not gaining ground and open weights buy ecosystem. This is an ecosystem bill, not an idealism bill.
Now, "local AI." Four Mac Studios at 46,999 yuan each comes to nearly 190,000 yuan, for 36.7 tokens per second. That is enough to write code, but it is several Macs away from "a local model for everyone." Apple is delighted, of course: the more local the model, the more Macs it sells. The people truly comforted are developers who want privacy and can afford a workstation. Everyone else will still choose between an API bill and an electricity bill.
One more line of fine print: open weights are not open source. The weights are there; the training data, code and method may not be, and the license keeps restrictions. Calling it "open source" is the marketing department's job; treating it as "free" is the finance department's disaster. Zuckerberg and Apple's new CEO John Ternus thanked each other, and the scene was warm. But remember: in this story of mom and dad finally getting along, the child is a model that needs four Macs to start walking.
Source: 36Kr · Vietnam News Agency