Records that name JoshuaSWarren as author. Nothing here is inferred from a screenshot.
Joshua Warren ships mlx-omarchy 0.5.0: first Core ML model runs end-to-end on Linux, Parakeet encoder on Apple Neural Engine matches macOS transcripts byte-for-byte, M1/M1 Max supported (M2+ WIP), with fewer Vulkan dispatches per token.
JoshuaSWarren reports Omarchy short-prefill now 33% faster than macOS Metal on the same M1 base, with long-context still catching up and broader Apple Neural Engine model support underway.
JoshuaSWarren demos Omarchy agents on an M1 MacBook Pro driving an M2 over USB to probe hardware details — agentic multi-Mac workflow.
JoshuaSWarren shares rough bring-up notes for an Apple Neural Engine driver targeting M2 under Omarchy, with a first alpha expected soon.
JoshuaSWarren asks owners of M1 Pro, M1 Ultra, base M2, M2 Pro, M2 Ultra, or any M3 or newer Mac to help bring the Apple Neural Engine up on Omarchy by running an updated collector under macOS and Omarchy, linking joshuaswarren/omarchy-mlx docs/contribute-data.md.
@JoshuaSWarren · M1 Max
JoshuaSWarren says Omarchy on Apple Silicon makes local AI fast with his MLX and Apple Neural Engine drivers, that dhh inspired him to believe computing should be fun, look good, and run fast, and that he built Coreglass to make inference profiling fun, fast, and beautiful, linking joshuaswarren/coreglass; the screenshot is a synthwave COREGLASS dash titled "Where the speed goes" for M1 Max · T6001 · Omarchy M+ · linux-aurora with levers such as 78 ms Vulkan submit length at 13.0x and a note that ANE is idle during LLM decode.
JoshuaSWarren says two more Mac models got their Neural Engine running on Omarchy today, with SiliconForested running the first M1 Pro test and ncfrontiersman the first M2 Pro, both bit-exact, and that a 27B model on an M2 Max now starts answering 25% sooner under Linux, asking which Mac is next.
JoshuaSWarren says splitting AI models across Macs on Omarchy Linux hit a wall: Thunderbolt networking received at a third of its send speed, so JoshuaSWarren found the bug in Linux's Thunderbolt network driver and fixed it, from 6.3 to 16.6 Gb/s, with the fix going to the Omarchy M kernel first, then upstream for every Linux distro.
JoshuaSWarren says "Apple says you need a Mac to build iPhone apps." Today they built one on Omarchy Linux, submitted it to the App Store, installed it on their iPhone over USB, and stopped it at a breakpoint in the debugger: "No Mac anywhere." They say it's all open source, linking joshuaswarren/omarchy-apple-dev. In a follow-up reply they say it is built on top of some awesome existing open source projects, that they are submitting PRs back to many of them that enabled the true end-to-end, Mac-free workflow, and that it runs on Omarchy running on x86 as well as Omarchy running on Apple Silicon.
JoshuaSWarren says "For those that have been asking when the M3 will be supported in Omarchy Linux, kernel support is here!" with M3 Pro in this release and other variants currently under testing, adding "Run Linux on your M1, M2, M3 today, including my ANE and GPU work for fast local AI." In a follow-up reply they say "I need volunteers for the M3 ANE testing before we enable the driver for everyone.", with first steps outlined in joshuaswarren/omarchy-mlx after getting the new kernel running on an M3.
JoshuaSWarren says "Your old Mac got a second life with Omarchy. Give it a second job in a cluster." GLM-5.3 is 753B params, 216 GB, and it ran as one endpoint across a Mac Studio, three MacBook Pros and a Dell laptop, on Omarchy: "Slow, but the full model runs - not a quant." They link joshuaswarren/omarchy-cluster, and in a follow-up reply say performance improvements to speed it up are coming next, as well as pulling in more interface methods, and eGPU support: "Right now it can use direct Thunderbolt links or the fastest network link you can give it, but I want to pull in @ashxhart's MCDMA and other methods too."
