Joshua Warren runs Meta’s 816B Muse Spark Open on three old Omarchy laptops and a Mac Studio

When Meta released the weights of Muse Spark Open, Mark Zuckerberg said on X on October 15, 2026 that the model “runs great on new Mac Studios thanks to work with @johnternus and Apple”. Joshua Warren, who works on Apple Silicon support as part of the Omarchy M team, answered with the hardware he already had. “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.

The run uses his own build of the model, Muse-Spark-Open-816B-A42B-MLX-2bit-mixed on Hugging Face. It is a mixed-precision MLX conversion of AMD’s FP8 release: experts at 2 bits, attention and embeddings at 4 bits, and the output head at 6 bits, for about 237 GB of weights under Meta’s Muse license. That is far more than any one of his machines can hold.

So the model gets split. According to the detailed report in the omarchy-cluster repo, all 62 layers ran on Apple Silicon GPUs through MLX, using omarchy-mlx on the Omarchy machines, with “no expert pruning, no 3-bit, no CPU-hosted layers, nothing substituted”. The work was spread over five processes on four machines:

• an M1 Ultra Mac Studio on macOS: 24 layers plus the output head (about 92 GB)
• a 14-inch M2 Max on Omarchy, running two processes of 12 layers each
• a 16-inch M1 Max on Omarchy: 12 layers
• a 13-inch M1 on Omarchy: 2 layers plus the embedding

Let’s be clear about what this is: a proof that it works, not a speed record. Warren measured decoding at about 2.0 to 2.2 tokens per second, with roughly 26 seconds before the first token and 10 minutes 48 seconds to load the model. It did answer correctly, naming Canberra as the capital of Australia and writing an is_prime function, and a repeat pass gave byte-identical output. The report is candid about where the time goes: about 45 ms per token on the Mac Studio, about 450 ms across the four laptop processes, “so the Linux ring path sets the speed”.

It wasn’t flawless either. The report notes that the 13-inch M1’s GPU hung once during a longer 128-token capture, with a Vulkan timeout, and the ring dropped. Peak memory and prefill speed haven’t been measured yet.

The bigger idea is omarchy-cluster itself, an MIT-licensed tool that serves one local model across Macs on macOS and Macs on Omarchy as a single OpenAI-compatible endpoint. Its README frames it as an extension of Omarchy’s promise that an old machine still has a second life: a MacBook too small for a modern model on its own can still hold some of its layers. The same repo also records a separate Muse Spark Open run on three machines through llama.cpp instead of MLX, using a smaller 1-bit GGUF.

Warren isn’t stopping there. Later on October 16, he said his new M6 Mac Mini had just been delivered and promised “a speedrun to see just how quickly I can get the new Apple Neural Engine in the M6 up and running under @Omarchy Linux”. No results yet, but we’ll be watching.

Source: https://x.com/JoshuaSWarren/status/2111084803781009688