One month, two developers: Linux gets a working GPU driver for the M4 Mac mini, running Minecraft at 200fps

Two developers have produced a fully OpenGL ES 3.0 compliant GPU driver for Apple's M4 chip in about a month, and it runs Minecraft at 200 frames per second 1. Cody Ho and Niklas published the work alongside their code, with driver commits running through September 15 2, and they also have Chrome and Firefox running WebGL on the M4 Mac mini with working compositing, on a Linux kernel driver they wrote for the GPU's firmware 1. The code is not ready for end users, and the 200fps figure is the pair's own benchmark. The number that matters more is the calendar: working OpenGL ES 3.0 took the volunteer Asahi Linux project roughly twenty-nine months from the January 2021 start of the work its first progress report covered, a gap of nearly thirty to one 34. That gap is the story, and it was not bought with easier hardware.

What "compliant" means here, and what it does not

Compliance is precise and self-attested: the Mesa driver passes the Khronos compatibility test suite for OpenGL ES 3.0, with unsupported tests limited to optional extensions 1. It is not the bar Asahi cleared in August 2023, when that project shipped drivers it called the world's only conformant OpenGL ES 3.1 implementation for M1- and M2-family hardware 5. What the M4 driver demonstrates today, by the authors' account: WebGL in two browsers, working compositing, Minecraft at 200fps, and a Mesa stack with a custom shader compiler and command stream builder 1. What it does not have, from the pair's own remaining-work list 1:

• Vulkan 1.4, OpenGL 4.6, OpenGL ES 3.2, and OpenCL 3.1
• Direct3D 12 via Proton, and ray tracing
• An end-user release of any kind

The reverse engineering also surfaced hardware the vendor's own stack ignores: a native single-instruction 64-bit add, and anisotropic filtering up to 128x, where Metal caps at 16x 1.

The first driver of this class took two and a half years

The Asahi project's own dated posts make the comparison measurable:

• March 2021: the first Asahi progress report covers January and February work on M1 support 5
• December 7, 2022: first public alpha Apple GPU driver, usable for a desktop and some games 5
• June 6, 2023: working OpenGL ES 3.0 reaches users, still experimental and not yet conformant 4
• August 22, 2023: first conformant drivers, OpenGL ES 3.1, for M1 and M2 5
• 2024: a single developer ships conformant Vulkan 1.3 on the M1, in what its author describes as one month of work 6
• September 13 to 15, 2026: the M4 kernel and Mesa driver commits land, the kernel leg taking three days by the authors' account 17

How the month happened: Ho had already built a hypervisor for tracing macOS, and the driver was reversed from those hardware traces alone, with no Apple binaries read 1. The Codex coding agent worked as the replay engine: at each firmware event it saved the GPU's entire memory state, restored it after a reboot, reconstructed the objects in code, then kept shrinking the replayed pages until everything was built from source 1. The firmware interface Ho reversed first, on the A18 Pro, has 1.5x as many structs and twice as many pointers as the M1 firmware interface that Asahi Lina decoded over what Ho calls grueling 12-hour days, and per the post's addendum the M4's firmware ABI differs significantly from the A18 Pro's 1. The pair also credit the M1/M2 work with making the problem's shape familiar: the exact semantics differ, but many of the right tradeoffs had already been made for them 1.

The agent side carries its own lesson. Claude could enumerate the new instruction set but stalled when asked to build working programs; the breakthrough came when Niklas grounded his agent in building Mesa itself, which spent its time more effectively than Ho's completeness-first exploration, by his own account 1. When Codex stalled under cybersecurity guardrails, Ho wrote a daemon that screenshotted the agent every minute and typed a resume command whenever the screen stopped changing 1. The month was not one brilliant prompt; it was a workflow that kept a tireless, easily distracted worker pointed at hardware.

A platform dependency owned by two volunteers

Nothing ships from this yet, and the road into mainline Linux is long. The pair want to upstream the Mesa driver but say it needs far more testing, human review, and refactoring, and they expect skepticism toward what they call "likely the first ever fully LLM-written GPU driver" 1. The kernel driver's path is worse: they will wait for the M1/M2 kernel driver to merge first, which they note may be a while 18. That wait is real. Meanwhile Asahi's official support reached the M3 series only on September 6, 2026 5, so the M4 work targets silicon beyond what the mainline project officially supports, carried by two people, their agents, and a Discord server 1.

The provenance is at least inspectable: the pair published their experiments in the twin agx-re repositories so anyone can verify the clean-room chain, and they discarded any memory page found to contain a shader, substituting their own, as a copyright measure 1.

For developers and fleets weighing M-series hardware for Linux, today's support matrix is unchanged. What changes is the planning assumption: new Apple silicon used to mean a years-long wait on a specific set of volunteers, and the M4 shows that wait compressing to weeks when someone arrives with a mature playbook and an agent to run the grind. The M1/M2 authors, whom Ho and Niklas call "utter wizards," paid the years that made the month possible 1. The open question is who pays for the next five years of maintenance, review, and trust, and no coding agent has answered that yet.