816B Open Weights on Two Mac Studios: Meta Open-Sources Muse Spark Open with Apple

On October 15, Meta open-sourced the 816-billion-parameter Muse Spark Open, billed as “the most capable American open-weights model to date”; the same day, Apple's machine learning team published an MLX blog demonstrating it running locally on two to four M5 Ultra Mac Studios. Zuckerberg publicly thanked Apple's new CEO John Ternus on X — old foes, suddenly sharing a frame.

The Qualifier in “Strongest” Is “American”

The wording is the tell. Meta's own comparison table concedes that on Terminal-Bench, DeepSWE long-horizon software engineering and GDPval professional-work evaluation, Muse Spark Open trails China's Kimi K3 and GLM 5.3; it leads mainly on tool-use tests like JobBench and MCP-Atlas. Meta Chief AI Officer Alexandr Wang was blunt: the strongest models usually stay closed for safety, but “the majority of AI in the world will run on open models — and those should come from America.”

How You Fit 816B Parameters into Two Macs

The bottleneck is memory: at native BF16, parameters alone exceed 1.6TB, while a maxed Mac Studio offers 512GB unified memory. Apple's two-step answer: NVFP4 four-bit quantization compresses weights to ~493GB, then Thunderbolt 5 direct links mesh multiple machines. Measured results (8K context, 128 generated tokens): two-machine pipeline parallelism at 15.74 tokens/s, two-machine tensor parallelism at 24.99, four-machine at 36.73, with first-token latency cut from 8.2s to 4.8s. In demos, the model fixed bugs and completed builds for an open-source iOS app entirely locally, and scrubbed footage frame by frame in DaVinci Resolve.

Closed Flagship Sells Subscriptions; Open Runner-Up Wins Ecosystem

It is Meta's classic double play: days after launching the closed Muse Wonder at $200/month (covered in this column on October 13), it open-sources the runner-up. For Meta, open weights are the time-tested weapon against OpenAI and Anthropic; for Apple, “the strongest American open model runs on two Macs” is the best possible ad before M5 Ultra Mac Studio sales begin — and the local-inference narrative grows straight out of Apple's privacy DNA.

The deeper signal is the migration of deployment rights: a frontier-class model now has an official reference implementation where data never leaves the device. For compliance-sensitive government and enterprise scenarios, open weights plus on-prem inference means the model may be open, but the boundary must be your own — the same logic Jiutian's Tiandun system applies in high-sensitivity deployments: running locally is the starting point; governing access is the license to operate.

(Compiled from 36Kr, Huxiu, OSCHINA and pedaily.cn public reports)