Takeaway: On November 7, 2026, Asahi Linux published its Linux 7.3 progress report, written by James Calligeros. The headline policy change is a license shift: the project is beginning to move its repositories from the permissive MIT license to the GNU General Public License version 2 or later. The stated reason is not a sudden dislike of sharing. Asahi says outsiders tried to exploit MIT’s terms to hijack in-progress work — including copying unfinished features from personal forks without attribution, or rewriting that code with LLM help to hide where it came from — with the aim of replacing the project and monetizing Linux on Apple Silicon. The same report also makes M3 support generally available, ships a limited first cut of Thunderbolt, and describes Windows-game experiments that the project is not ready to ship.
This is an open-source licensing and desktop-Linux story. It is not the same topic as our earlier note on an AI pipeline that found Linux kernel local privilege escalations. That piece was about vulnerability discovery. This one is about how a reverse-engineering project thinks it has to protect its own code — and what users of Fedora Asahi Remix actually get in the 7.3 cycle.
Why Asahi used MIT in the first place
Asahi Linux reverse-engineers Apple Silicon Macs so they can run a real Linux desktop. When the project started, it deliberately licensed as much of its own code as possible under MIT. The goal, the report says, was to make it easy for “friends at the BSDs and other Free Software operating systems” to reuse the work. A permissive license is a good fit for that kind of cross-project sharing: other kernels and userlands can take the code without being forced onto the GPL.
That choice always had a trade-off. MIT requires that copyright and permission notices travel with redistributed copies, but it does not require someone who ships a modified version to publish their own source. Asahi now says that gap is no longer acceptable for its repositories.
What the project says happened
The report is unusually direct about motive. Calligeros writes that practices last common in the 1990s and 2000s — treating permissively licensed projects as “free real estate” — have returned, “driven in large part by both the insatiable appetite of frontier AI labs for high-quality training data, and militant animosity toward the Free Software community.”
The concrete incidents Asahi describes are about its own trees, not a general complaint about model training:
- Multiple attempts to exploit the MIT license in order to hijack Asahi’s work.
- A stated aim, in Asahi’s telling, of replacing the project and “aggressively monetising Linux on Apple Silicon.”
- In extreme cases, in-development feature code lifted from developers’ personal forks and published without attribution.
- Or the same code “lazily rewritten with LLM assistance to obfuscate its origin.”
Asahi’s legal point is narrow and important: “The MIT licence explicitly allows this, and has no carveouts to protect against bad faith actors.” The project is not claiming those copies were secret copyright infringement under MIT. It is saying MIT’s permissions were used in ways that threaten the project’s survival, so the license has to change going forward.
As such, we made the difficult decision to begin the process of transitioning all of our repositories to the GNU General Public License version 2 or later. Going forward all Asahi Linux code will be licenced under strong copyleft licences only. Those who do not learn from history are doomed to repeat it.
James Calligeros, Asahi Linux, November 7, 2026
Two phrases in that statement matter for anyone who already pulled Asahi code. First, this is “the process of transitioning” all repositories — not a claim that every tree flipped on the same day the report went up. Second, “going forward” points at new licensing of Asahi’s own code. Copies that were already distributed under MIT generally keep the terms they were received under; a later relicensing does not reach back and rewrite those older grants. What changes is what new work, and newly relicensed repositories, require of people who distribute derivatives.
The Wine precedent — and what GPLv2 actually requires
Asahi compares the move to Wine. Wine originally shipped under MIT. In the 1990s and 2000s, Asahi writes, companies built monetized proprietary forks, which was allowed. Wine’s developers worried those forks would fragment the ecosystem until Wine itself became unviable. In 2002, Wine switched to the LGPL so later proprietary forks would be harder to sustain and improvements would be more likely to flow back upstream.
Asahi is choosing a stronger copyleft than Wine did: GPLv2-or-later, not LGPL. Under the GPL, anyone who distributes a derivative work to other people must pass along the same freedoms, including corresponding source for binaries (or another step the license permits, such as a qualifying written offer). The GPL does not ban commercial use or sale. A company can still ship Asahi-derived code for money. What it cannot do, once the code is under the GPL and it distributes a modified version, is keep the corresponding source private from the people who received the binary.
The report also frames the change as a values decision, not only a defensive one. Asahi says it develops in the open — finances, branches, progress reports, IRC, and Matrix — because it treats Free Software as an end, not only a convenient way to get Linux onto Macs. LWN noted the license section the same day, and the project’s own Mastodon account highlighted it as well.
M3 support is now generally available
The license news is sitting on top of a substantial hardware cycle. Asahi previously offered initial M3-series support only behind the installer’s EXPERT flag. The project says early adopters filed almost no issues, so M3 support is now generally available without that flag.
The same update enables DCP on M3 machines. Asahi says that brings performance and battery-life improvements, makes sleep work on M3-series devices, and makes the area around the MacBook notch usable. Users are told to watch for m1n1 and kernel updates in the days after the report.
GPU support is the obvious remaining gap versus M1 and M2. Asahi describes Apple’s GPU cadence as a tick-tock, with the M3 GPU a large break from earlier designs. At XDC this year, Alyssa Millburn, Anna Maniscalco, and TellowKrinkle presented reverse-engineering progress far enough for m1n1 to initialize the GPU and draw a triangle. The report warns that Apple’s design choices will take significant work to integrate into upstream Mesa. That is a research milestone, not a claim that M3 GPU acceleration is ready for daily desktop use.
Thunderbolt is shipping — with a sharp caveat
Asahi promised Thunderbolt support “within the year” back in January 2024 and then said almost nothing about it. The 7.3 report is the update: the patchset went upstream first, and the downstream kernel is now shipping it to users.
Apple designed its own Thunderbolt controller rather than using Intel’s widely available parts. Asahi describes each Thunderbolt-capable USB-C port as a Cortex-M3 running RTKit plus a custom Thunderbolt/USB4 router. Linux’s Thunderbolt subsystem, by contrast, was shaped around Intel controllers and their assumed topology, register access, and shared memory. Apple’s design also ties each controller’s reset line to bits in the middle of PMGR’s MMIO range — a pattern Asahi says no other block uses quite the same way. Sven’s work on that mismatch is what finally produced a patchset.
What users can do today is limited on purpose:
- Supported now: USB3, XDomain, and Thunderbolt Networking — you can attach USB devices, or connect two Thunderbolt computers and ping each other.
- Not supported yet: PCIe and DisplayPort tunneling, which most Thunderbolt docks need for storage, Ethernet dongles beyond simple USB, and extra displays.
If you buy a dock expecting it to behave like it does in macOS, this first drop will not deliver that. Asahi is explicit: “That’s about it” beyond USB3, XDomain, and networking, and “Watch this space” for the rest.
M4, M5, and three virtual machines under m1n1
Upstreaming continues in parallel. Two large drivers, macaudio and AVD, went onto the list this cycle — about 3,100 lines for the audio work and related Devicetree changes, and about 13,000 lines for AVD, which spent barely three months in Asahi’s downstream tree. Initial Devicetrees for M4 Pro and Max machines, and for M5 thanks to contributor Yureka, are also on the list. Those trees are bare-bones: enough to boot the SoC to a TTY, not a full desktop. Some work, including changes to the TI USB-PD and SPMI drivers and the initial Thunderbolt support, was developed upstream-first.
Reverse engineering newer chips is getting harder. Asahi had already teased virtualizing SPTM-enabled macOS in m1n1 so developers could inspect M4-series machines. M5 adds Exclaves: small applications that run in an isolated environment controlled by the Secure Kernel. On M5 and later, macOS uses them for sensitive jobs such as drawing the “camera in use” indicator. XNU talks to the Secure Kernel and its Exclaves through SPTM, using another IPC protocol called Tightbeam, and panics if those Exclaves are missing. Asahi compares the shape of this design to Arm TrustZone.
Some Exclaves can still be bypassed with command-line settings, but Asahi does not expect that to last. Sven tried putting the Secure Kernel and Exclave userspace in another VM beside SPTM and macOS — “Oh, that worked” — and m1n1 can now spawn three VMs in EL1. The practical result the report claims is the ability to virtualize and reverse-engineer macOS on both M4 and M5 machines.
Games outside muvm: a demo, not a release
Playing retail Windows games on Asahi has required a stack of translators: FEX to turn x86/amd64 binaries into arm64, Wine (now with mature ARM64EC support) for the Windows API, and muvm, a thin VM that presents the 4 KiB page size many games assume. Asahi says FEX and Wine have improved a great deal, but muvm’s performance penalty is still significant — about 20%, depending on the workload — and the VM cannot use Wayland, has slow I/O, and struggles with some input devices.
macOS has an advantage here: XNU can map userspace at mixed page granularities, so Rosetta can ask for 4 KiB pages for a child process. Linux maps every process at the kernel’s page size, which on Apple Silicon Asahi is 16 KiB. A previous hack inside FEX that let some Windows games run outside muvm was removed to keep translation correct.
The new experiment uses P3S (Per Process Page Size), work Google engineers Kalesh Singh and Frederick Mayle described at Linux Plumbers Conference for Android’s move from 4 KiB to 16 KiB pages. P3S tracks each process’s page-table granularity and writes the Translation Control Registers when the scheduler switches to a process that needs a different size. Asahi says the extra work is effectively a single register write, and a no-op for processes that already use the native size.
Rebased onto the Asahi tree, that patch let the team run several titles outside muvm, including Star Wars Battlefront (2004) with WineD3D, Wolfenstein: The New Order (2014), Grand Theft Auto IV (2008) with DXVK, and Fallout: New Vegas (2010) with DXVK. Wolfenstein hit more than 144 frames per second on an M2 Max Mac Studio. GTA IV was “not fantastic,” but still much better than under muvm.
Asahi is nevertheless refusing to ship this in its downstream kernel. The current P3S code is hard-coded for 16 KiB hosts and 4 KiB guests, which matches Android’s immediate need but is not, in Asahi’s view, appropriate to merge as pagesize-specific kludges in core memory-management code. The project does not know whether Google will generalize and upstream the work. It also needs a userspace shim, because modern arm64 binaries are often aligned to 64 KiB and would not trigger P3S on their own. Until that uncertainty clears, Asahi does not want users to depend on an unofficial page-size path.
Smaller quality-of-life notes
The MacBook microphone beamformer, Triforce, is written in Rust. After the team replaced an FFT-based analytic transform with IIR filters, the compiler stopped autovectorizing the hot path. Linkmauve added SIMD-optimized code for arm64 and amd64, and Frédéric Bour contributed other optimizations. Asahi says Triforce is now almost twice as fast while using less CPU time.
Fedora Linux 45 was released on November 3, 2026, and Fedora Asahi Remix shipped the same day — the first time the Remix has landed on upstream Fedora’s release day.
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| Asahi published a Linux 7.3 progress report on November 7, 2026 | Confirmed | Authored by James Calligeros on asahilinux.org |
| The project is moving its repositories from MIT to GPLv2-or-later | Confirmed as a process underway | “Begin the process”; not a claim that every repo flipped that day |
| Reason given: attempts to hijack work, including unattributed copies and LLM rewrites of in-development code | Stated by Asahi | The report does not name the parties |
| MIT already forbade those copies | No | Asahi says MIT expressly allows them if its notice conditions are met |
| Old MIT downloads are retroactively converted to GPL | Not what a relicensing does | Earlier distributions keep the license they were received under |
| The GPL bans selling or commercially using the code | No | It requires source-sharing when derivatives are distributed |
| M3 support is generally available without the EXPERT installer flag | Confirmed by the report | DCP, sleep, and notch-area use are called out; GPU is not at M1/M2 parity |
| Thunderbolt docks work like they do on macOS | No | USB3, XDomain, and networking only; no PCIe or DisplayPort tunneling yet |
| Windows games now run outside muvm on shipping Asahi kernels | No | P3S demos are described; the patch is not being shipped |
| This report is about AI finding Linux kernel bugs | No | Different story from OpenAI’s kernel-LPE pipeline write-up |
FAQ
Did Asahi relicense every file on November 7?
No. The report says the project has begun transitioning all of its repositories to GPLv2-or-later and that, going forward, Asahi Linux code will use strong copyleft licenses only. Check each repository’s current license file for the date a given tree actually changed.
Can companies still use Asahi’s code in a product?
Yes, including commercially. The GPL allows use and sale. If you distribute a GPL-covered derivative to someone else, you must also give them the corresponding source and the same license rights. Using the code only internally does not, by itself, create a public-release obligation.
Why not stay on MIT and just ask people to behave?
Asahi’s argument is that MIT already permits the behavior it objects to — publishing modified copies, including LLM-rewritten ones, without sharing later improvements — as long as notices are preserved. A request for good manners is not a license term. Copyleft changes the terms for new distributions.
Does this help BSD projects that reused Asahi’s MIT code?
Older MIT snapshots they already received remain usable under MIT. New Asahi code placed under GPLv2-or-later is harder to import into a permissively licensed BSD tree. That is the trade-off Asahi originally tried to avoid, and the one it now says is necessary.
Can I use a Thunderbolt dock on Asahi after this update?
Only for what the first patchset supports: USB3 devices, XDomain, and Thunderbolt networking between two computers. Most docks need PCIe and DisplayPort tunneling, which Asahi says are not supported yet.
Will GTA IV or other Windows games just work now?
Not as a supported feature. The report shows titles running outside muvm with Google’s P3S patches applied locally. Asahi is not shipping that kernel change until the upstream story is clearer, and it says a userspace shim would also be required.
Sources: Asahi Linux — Progress Report: Linux 7.3 (James Calligeros, November 7, 2026); LWN.net — Asahi Linux 7.3 progress report (November 7, 2026). Quotations and technical figures above are taken from the official report. We did not independently verify the unattributed-copy incidents Asahi describes, and the report does not name the parties involved.