AV1 video shows shifted colours or blocky corruption on AMD graphics (linux-firmware-amdgpu 20260810-1)
Checked the Arch package archive with curl and a named user agent: 20260810-2 is dated 14-Aug-2026 15:52 and 20260810-3 is dated 08-Sep-2026 21:07, both as the record says, and a fourth build, 20260910-1, landed on 10-Sep-2026 which the record does not know about. The Arch packaging commits say what each build does, and that is the main defect: ba3cc99c is "20260810-2: Backport an upstream revert and revert all amdgpu VCN firmware" while cc4c7047 is "20260810-3: Remove VCN revert now that Mesa is fixed", so the record's "anything from -2 up carries the fix" is wrong. From -3 onwards the firmware revert is gone and the fix rests entirely on Mesa 26.1.8 or newer, so a machine on -3 or 20260910-1 with an older Mesa is back in the broken pair, and the fix and verify text had to be rewritten around that. Confirmed on this machine: /var/log/pacman.log records linux-firmware-amdgpu 20260810-1 installed at 2026-08-23T00:40:33+0000 and upgraded to 20260810-2 at 2026-08-22T17:52:45-0700, in the same transaction that took mesa from 1:26.1.7-1 to 1:26.1.8-1, which confirms the mirror claim as written. Mesa merge request 43787 is titled "radeonsi,radv/video: Fix AV1 decode qmatrix params", not "radv/video:" as the record has it, and it carries two commits, so the record quoted the Vulkan one while its own mechanism is VA-API, which runs through the radeonsi commit. On Q2 the answer is do not split the RDNA3 reports out: I read the radeonsi diff, src/gallium/drivers/radeonsi/mm/si_video_dec.c in si_dec_av1, and it is gated on using_qmatrix rather than on any ASIC, the Mesa 26.1.8 notes of 2026-08-19 list both commits, and the -2 package reverted all amdgpu VCN blobs rather than only the Navi 48 one, so one firmware regression covers both generations. Read both cited issues in full: 7377 supports the 780M and 760M reports and the Firefox workaround, and its 780M reporter confirmed -2 cleared it while the 760M reporter only ever confirmed the Firefox workaround, so I narrowed that sentence. 7514 does not support "after an upgrade from Omarchy 3.8.4": that reporter was running Omarchy 3.8.4 and read the broken firmware off the same Omarchy stable mirror, and his core.db last-modified of Fri, 14 Aug 2026 15:15:59 GMT sits 37 minutes before -2 reached the archive, which is better evidence for the lag claim than the record gives, so I put it in. Not exercised: this workstation has an NVIDIA card, so no AV1 decode on AMD hardware was reproduced here and nothing about the corruption itself was seen first hand.
Symptom
After updating, or on a fresh Omarchy 4.0.0-1 install, AV1 video in Firefox, Chromium and Brave shows rectangular areas in the wrong colours, sometimes pixelated. Lower resolutions such as 480p look fine and higher AV1 streams do not, for example YouTube av01.0.12M.08 at 4K while av01.0.04M.08 at 480p is clean. The corruption is in the decoded frames, so it appears in screenshots too.
Reported on a Ryzen 7 8745HS with Radeon 780M and a Ryzen 5 7640U with Radeon 760M, both on Omarchy 4.0.0-1, and on two Radeon RX 9070 XT machines, one of which was still on Omarchy 3.8.4 and read the broken firmware from the same Omarchy stable mirror. pacman -Q linux-firmware-amdgpu shows 20260810-1. The RX 9070 XT reports add Mesa 26.1.7-1, libva 2.24.1-1 and kernel 7.1.8-arch1-3.
Cause
linux-firmware-amdgpu 20260810-1 carries an AMD GPU firmware regression that corrupts hardware AV1 decoding. The browsers all reach the decoder through VA-API, which on Mesa is the radeonsi gallium media path, so every browser was affected.
The firmware changed how the AV1 decoder reads the quantizer matrix parameters. The high 4 bits of qm_y, qm_u and qm_v stopped being ignored, while Mesa was still OR-ing 0xf0 into all three unconditionally. Two independent fixes exist upstream and either one is enough, but they do not stack the way the version numbers suggest.
Arch published linux-firmware-amdgpu 20260810-2 on 2026-08-14. Its packaging commit ba3cc99c is "20260810-2: Backport an upstream revert and revert all amdgpu VCN firmware", so it reverts every amdgpu VCN blob rather than only the RDNA4 one.
Mesa merge request 43787, radeonsi,radv/video: Fix AV1 decode qmatrix params, merged on 2026-08-17 and shipped in Mesa 26.1.8 on 2026-08-19. It carries two commits, radeonsi/mm: Fix AV1 decode qmatrix params for the VA-API path the browsers use and radv/video: Fix AV1 decode qmatrix params for Vulkan video. The radeonsi change, in src/gallium/drivers/radeonsi/mm/si_video_dec.c, passes the raw quantizer matrix values when the stream uses one and 0xff when it does not, and it is not gated on any ASIC, so it covers every VCN generation.
Arch then published 20260810-3 on 2026-09-08 with packaging commit cc4c7047, "20260810-3: Remove VCN revert now that Mesa is fixed", and 20260910-1 on 2026-09-10 on top of that. From -3 onwards the firmware revert is gone and the fix rests on Mesa alone, so a machine on -3 or newer whose Mesa is older than 26.1.8 is back in the broken combination.
The thread established the mechanism on RDNA4 (gfx12, Navi 48), and the Radeon 780M and 760M reports belong in the same record rather than a separate one: they carry the same firmware version, the Mesa fix is ASIC-independent, and the -2 revert covered all VCN generations. The 780M reporter confirmed that -2 cleared it. The 760M reporter confirmed only the Firefox workaround and never reported back on the firmware update.
Omarchy's stable mirror lagged, which is why this reached people a week after Arch had fixed it. The reporter of issue 7514 read the mirror's core.db as last modified Fri, 14 Aug 2026 15:15:59 GMT, and -2 reached the Arch archive at 15:52 the same day, so the snapshot was taken about 37 minutes before the fixed build existed. The mirror was serving 20260810-2 by 2026-08-22, read from /var/log/pacman.log on an Omarchy stable workstation.
Fix
Update. Omarchy 4:
omarchy update
pacman -Q linux-firmware-amdgpu mesa
Plain Arch:
sudo pacman -Syu
Either upstream fix alone is enough, but check which one you ended up with, because Arch dropped the firmware revert again:
20260810-2carries the firmware revert and fixes this on any Mesa.20260810-3(2026-09-08) and20260910-1(2026-09-10) dropped that revert, so on those you need Mesa 26.1.8 or newer.
So the good pairs are 20260810-2 with any Mesa, or -3 and newer with Mesa 26.1.8 or newer. A full omarchy update gives you both halves and is the supported path. Do not pull the firmware package on its own, which leaves you with a partially upgraded system.
Reboot afterwards. GPU firmware is loaded when the amdgpu module initialises, so the running session keeps the old firmware until then.
If you cannot update yet, turn off the hardware AV1 path in the browser. One reporter on a Radeon 760M confirmed that disabling AV1 in Firefox, media.av1.enabled set to false in about:config, stops the corruption, because YouTube then falls back to VP9. For Chromium and Brave, launch with --disable-accelerated-video-decode, which the RX 9070 XT reporter used to isolate the fault to hardware decode. That same reporter found that disabling hardware decoding in Firefox alone did not help on his machine, so treat the Firefox route as unconfirmed there. All of this is a workaround, not the fix.
Verify
pacman -Q linux-firmware-amdgpu mesa
You want either 20260810-2 with any Mesa, or 20260810-3 and newer together with Mesa 26.1.8 or newer. Arch's mesa carries an epoch, so 26.1.8 reads as 1:26.1.8-1.
After a reboot, play a 1080p or 4K AV1 stream with hardware decoding on: colours are stable and no blocks appear. One reporter on Omarchy 4.0.0-1 confirmed the firmware update alone fixed it, and another confirmed Mesa 26.1.8 alone fixed it with the firmware still on 20260810-1.
Sources
- https://github.com/omacom/omarchy/issues/7377
- https://github.com/omacom/omarchy/issues/7514
- https://archlinux.org/packages/core/any/linux-firmware-amdgpu/
- https://archive.archlinux.org/packages/l/linux-firmware-amdgpu/
- https://archlinux.org/packages/extra/x86_64/mesa/
- https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/43787
- https://gitlab.archlinux.org/archlinux/packaging/packages/linux-firmware/-/commits/main
- https://gitlab.freedesktop.org/mesa/mesa/-/commit/3c4d3e46
- https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/docs/relnotes/26.1.8.rst