Joshua Warren says his Mesa work makes local AI 18x faster on an M1 Max running Omarchy
Joshua Warren has opened a stack of pull requests against Omacom’s Mesa fork that tune the Honeykrisp Vulkan driver for Apple GPUs, and he reports a jump from 1.5 to 27.4 tokens per second on an M1 Max.
Running AI models on a Mac’s GPU under Linux has been slow. Joshua Warren decided to fix that in the driver itself. In a post on X on October 8, 2026, he wrote that he “went into Mesa, the open source GPU driver, and made it 18x faster.”
His numbers, measured on an M1 Max running Omarchy: decode went from 1.5 to 27.4 tokens per second, and prompt processing went from 55 to 237 tokens per second.
The code is in pull request #6 on omacom/mesa, titled “honeykrisp: M1 compute speedups and the deferred CDM barrier on G13G.” It is the top of a stack and builds on pull requests #3, #4 and #5. At the time of writing, the pull request was still open and had not been merged.
The 18x figure is Warren’s headline for the whole effort. PR #6 itself lists smaller, step-by-step gains on M1, each checked against identical token digests. Paired matrix loads add 9.4% to prefill, folding shared-memory offsets adds 1.8%, skipping the register-pressure scheduler for compute shaders adds 1.4% to 1.8% to decode, and deferring the CDM barrier adds another 0.9% to 1.4%.
The new defaults depend on the chip. M1 (G13G) and M1 Max (G13X) get the faster paths on by default. On M2 (G14), cooperative matrices stay off unless you opt in with AGX_SIMDMAT=1. Each change can also be switched off with an environment variable.
Warren works on GitHub as joshuaswarren, the same account behind omarchy-apple-dev, the project that builds and ships iOS apps from Omarchy without a Mac. This is another sign that Omarchy on Apple Silicon is turning into a serious machine for developers, not just a curiosity.
