A new Linux distribution for Apple Silicon has done something the long-established Asahi project has not: boot a GPU-accelerated desktop on the M4 Mac mini. And it credits AI coding agents for the speed.
Gravity Linux describes itself as a Fedora remix that brings a modern Linux desktop to Apple Silicon. Its About page states the project was “initially forked from Asahi Linux over differences in policy, namely regarding LLM use,” while emphasising respect for the older project and a shared goal.
## What the alpha actually does The first experimental alpha targets the M4 Mac mini, a machine Apple launched in October 2024. It ships GPU acceleration with OpenGL ES 3.0 and OpenGL 3.3 support and is explicitly aimed at developers rather than daily use. HDMI output works; USB-C displays, USB4, Thunderbolt and suspend do not, and power management remains incomplete.
Even so, that puts Gravity ahead of Asahi on hardware generations. Asahi has only recently added preliminary M3 support, without graphics acceleration, while Apple has already moved on to M6-based Macs. The comparison is not exact: Gravity's M4 build is an early developer alpha, whereas Asahi is a longer-established project covering far more hardware.
## Why the method is the controversy Gravity's lead, Cody Ho, previously worked at both Apple and OpenAI. With co-developer Niklas Sheth he published a two-part account of AI-assisted Apple Silicon reverse engineering, “I Came, I Prompted, I Left” (3 August and 15 September), describing first a custom hypervisor for newer Apple Silicon and then an M4 GPU driver built in about a month. Ho says much of the work ran through largely unattended LLM loops.
This is the sharpest difference from Asahi, whose Generative AI policy “broadly forbids the use of generative AI tooling for material contributions.” Gravity permits such tools but tries to contain the legal risk through procedural separation. Its LLM Use and Clean-Room Policy states that anyone whose session has looked at disassembly, decompilation or other protected implementation details of an Apple component must consider themselves “tainted” with respect to that component.
The rule was applied in practice. While developing the hypervisor, Ho used an LLM to disassemble Apple's proprietary SPTM component, which made him ineligible to write the corresponding clean-room implementation; the knowledge was documented and handed to someone who had not seen the protected material. The GPU work took a stricter route — hardware traces, live probing and self-written shaders rather than inspecting Apple binaries — while leaning on Asahi's earlier M1 and M2 work and its userspace API.
## What it does not prove Gravity's developers are careful about the claim. The code still needs extensive testing, human review and refactoring before it can be upstreamed, and the kernel driver may have to wait for Asahi's M1 and M2 driver to land upstream first. The release is therefore an intriguing demonstration of what AI-assisted reverse engineering can do — not a controlled experiment proving that Asahi's policy was the bottleneck.
