Update to my earlier post, Meta’s Muse Spark Open: 816B Open Weights, Two Mac Studios. When Meta released the weights, Mark Zuckerberg said the model “runs great on new Mac Studios.” Joshua Warren of the Omarchy M team answered with the gear he already had.
What he did
“I ran all 816B parameters (2-bit) on gear I had: 3 old M1/M2 laptops on @omarchy + a 2022 Mac Studio on macOS. No $40K cluster needed,” he posted on October 16.
- His own mixed-precision MLX build: experts at 2-bit, attention and embeddings at 4-bit, output head at 6-bit, about 237 GB of weights.
- All 62 layers on Apple Silicon GPUs via MLX, with no expert pruning or CPU-hosted layers, across five processes on four machines: an M1 Ultra Mac Studio (24 layers plus the head), a 14-inch M2 Max running two processes, a 16-inch M1 Max and a 13-inch M1.
- Speed: about 2.0-2.2 tokens/sec, ~26 seconds to first token, almost 11 minutes to load. Answers were correct and a repeat pass was byte-identical.
- Not flawless: the M1’s GPU hung once on a longer run with a Vulkan timeout.
The bigger piece is omarchy-cluster, an MIT-licensed tool that serves one local model across Macs on macOS and Macs on Omarchy as a single OpenAI-compatible endpoint. Warren also says his new M6 Mac Mini just arrived and he’s going after its Neural Engine under Omarchy next.
My take
This is a proof it works, not a speed record, and that’s the point. Old laptops sitting in a drawer can each hold a slice of a frontier-size open model. That’s local AI for regular people.
