The Apple Neural Engine is coming up on Omarchy
An Omarchy M team member bringing the Apple Neural Engine up on Omarchy asked owners of later Apple chips to run a data collector, in this request.
Joshua Warren asked for data from the M1 Pro, M1 Ultra, the M2 family, and any M3 or newer Mac, collected under both macOS and Omarchy. The instructions are in his omarchy-mlx repository.
David Heinemeier Hansson reposted the request.
Warren said the team is running its own models and open-source models on the engine. The M1 MacBook Air and the M1 MacBook Pro use the same Neural Engine, so the M1 data is already in, he said.
Warren said he has run a 2B Qwen model through the engine under Linux, and that Parakeet's speech encoder runs entirely on it. The engine works in fp16, so a 2B model takes about 4 GB.
On M1 chips the engine can address 4 GB at a time, so larger models have to be split into stages, Warren said. The M2 can address much more.
Asked whether he had enough M1 Max data, Warren said he thought so, and that the next wave would need people willing to test a beta driver.
Anyone who cannot open the link can find the instructions through the README of the omarchy-mlx repository on GitHub, Warren said.
Correction
Named Warren where the article said “they,” reported DHH's repost, which the sources already listed, linked the data collector instructions, and relabelled a source that pointed at a post rather than a repository.