Changelog

Booted on the Legion 5 15IAX10 (RTX 5070 Laptop, macOS 26.7.1 25G241, BIOS Discrete) on 2026-10-27 (N-294): Metal 3, 1680x1050 HiDPI at 165 Hz. System Settings lists the scaled sizes, the Dock blur has no blocks, the MPSGraph depthwise test gives the right values through the installed plugin, and window drag plus typing in an OpenGL terminal showed no text jumping back. The plugin on that boot was the xe-2.6.2 build, which contains every plugin change below.

• Rebased onto NullMoth 2.6.2, still on NVIDIA 615.78.08. The fork's commits from 2.5.3-nvgfx.1 and 2.5.3-nvgfx.2 (async cache writes) and the Intel panel proxy head (nvfbigpu=1, off by default) carried over with no code conflicts. NVRM rebuilt for 615.78.08 matches the shipped 2.6.2 NVRM in 30875 of 30884 functions, as on 2.5.3; NVAccel is upstream's, unchanged.
• NVRMFB: scaled modes carry kDisplayModeValidForHiResFlag, so HiDPI "looks like" sizes reach the public mode list (upstream #354). Built from source, the rest of NVRMFB is identical to the shipped 2.6.2 one.
• Plugin: when a shader brings its own sampler (a constexpr sampler, or one read from an argument buffer), the translator parks it in a sampler slot the shader leaves free. The app's sampler in that slot is now put back as soon as a shader no longer needs it, as Metal requires. Core Animation binds a linear sampler for the Dock's blur upscale once, and the upscale read a parked nearest sampler instead and drew 32x32 blocks (upstream #363, diagnosed by ccyykk666; our I-049). NVMTL_NO_SAMPLER_RESTORE=1 keeps the old behaviour.
• Translator: GeekAudio's PR #385 (immutable local pointer fields forwarded after a dominating store). With it MPSGraph's depthwiseConv_3x3 translates and computes the right result on the RTX 5070; with the shipped translator the same test aborts in MPSKernel MTLComputePipelineStateCache (Geekbench 6's CPU run hits this, our I-117). The new pass only runs on a module whose first translation failed. All translator test targets that build from the published tree give the same results as 2.6.2's, plus the PR's 11 new tests.
• Plugin: NVMTL_NO_PSO_SKIP=1 and NVMTL_NO_SV_MEMO=1 (nvrm610.conf) turn off the skip of a repeated setRenderPipelineState and the sampler-variant memo, to A/B a text glitch reported in Alacritty against them.

• Plugin: shader cache files are written, and their last-use time touched, on one background queue instead of on the thread that asked for the shader. WindowServer translated shaders and paid an fsync for each stored file on its coreanimation.pipelines thread, which fits window drags at 11 to 28 flips/s for the first minutes after login (upstream #370). Reads and the in-memory memo are unchanged, so a process still never reads or translates the same shader twice. NVMTL_CACHE_SYNC_IO=1 (environment, or a line in nvrm610.conf) puts the writes back on the caller for an A/B. Each cache file's last-use time is touched once per process, and a process waits up to 2 s at exit for its queued writes.
• Plugin: the log now names the shader cache folder in use and, after 500 function-constant lookups (then every 2000), how many were served from memory, from disk, or translated.

• Rebased onto NullMoth 2.5.3, which moves NVRM to NVIDIA's 615.78.08 driver and GSP firmware. Upstream now carries three things this fork used to: the boot hold that waits for the bring-up (our PR #317, nvrmholdcap), the function-table pipeline link (#318) and desktop scaling at the head viewport (#273); the fork's copies are dropped.
• NVRM rebuilt from source for 615 with the fork's power code (115 W TGP on AC, NVPCF and GPS _DSM, physctl): build_nvrm.sh takes NVRM_TAG (default 615.78.08), drops NVKMS's -mfunction-return=thunk-extern (XNU exports no __x86_return_thunk) and applies 615-viewport-scaling.patch, which upstream's patch notes list but its published completion script leaves out. The result matches the shipped 2.5.3 NVRM in 30875 of 30884 functions.
• Plugin: upstream's host-import surface upload (#296) and the fork's asynchronous upload at bind live side by side; an A/B is still to do.
• Booted on the Legion 5 15IAX10 (RTX 5070 Laptop, macOS 26.7.1, BIOS Discrete): Metal 3, 1680x1050 HiDPI at 165 Hz, 115 W on AC, GSP bring-up 13 s (40 to 48 s on 610).

• NVRM: the boot hold no longer lets WindowServer in while the card is still coming up. The cap used to fire at settle + 40 s whatever the bring-up was doing; on a Legion 5 15IAX10 the GSP bring-up takes 2 to 48 s from one boot to the next, and every boot over 40 s ended on software rendering (WindowServer asserted in CGXMappedDisplayStart, the driver then refused to arm). The cap now waits in 10 s steps while go(2) runs, up to nvrmholdcap=<s> (40 to 600, default 120), and the log says how long go(2) took.
• NVRM: the 115 W TGP mode follows the charger again (the 2.4.1 base caches the power source; the TGP tick refreshes it before reading).

• The same driver as 2.6.0, released with fixes to the 1401 apps for Mac and Windows.

• The same driver as 2.6.0, released with the redesigned 1401 apps for Mac and Windows.

• An update of the driver takes one restart. macOS shows a notice that the extensions were updated while the installer replaces them, and the installer used to answer it with "allow the extensions, then reboot once more"; the new driver already loads on the first restart (measured: every extension approved, none held back), so on an update it now says one restart is enough. A first install still asks you to allow the extensions once.

• Discord videos play again. 2.5.4 stopped registering the hardware video decoder in every sandboxed process, and Discord decodes inside its own sandboxed GPU process, so its videos did not play at all. Only Apple's sandboxed apps (Safari, QuickTime, Preview) now hand decoding to macOS's video decoding service; Chromium and Electron apps decode in their own process as in 2.5.3.
• Remote-desktop apps such as UU Remote show the picture: the hardware H.264 and HEVC encoders marked every keyframe "not a sync frame: false" where Apple's encoders leave the mark out, and apps that look for the mark's absence never found a keyframe on every card generation (#388, #358, thanks a1461750564 and sherves for measuring it). The encoders for the newest cards also load on macOS 15.0 to 15.4 now; they asked for 15.5.
• Screen Sharing and other screen-capture apps show the screen instead of black: a texture made from a buffer could only be viewed as its own format and its sRGB twin, and screen capture views a BGRA surface as RGBA when it scales or converts.
• Brave and Microsoft Edge get Metal through ANGLE the way Chrome does (PR #381, thanks SameerAlSahab; Edge tested by sherves).
• On a card with a small BAR1 window, NVAccel no longer logs every refused allocation: a third of all log lines on an idle desktop (#387, thanks wwwljxw). It logs the first eight and then every 256th with the running count, as the framebuffer does; debug.nvaccel_vm_refused keeps the exact total.

• Desktops whose only graphics card is the NVIDIA card no longer loop on a black screen when the card's display does not come up: WindowServer used to stop at start with "Failed to create MetalDevice" every ten seconds until the startup guard switched the driver off. It now starts on the firmware's framebuffer, the desktop appears without acceleration, and the first-login report sends the logs that show why the display did not come up.
• Discord gets Metal through ANGLE the way Chrome does (PR #348, thanks lukeitsback).
• Safari, QuickTime, Preview and other sandboxed apps decode H.264 and HEVC through macOS's video decoding service instead of failing inside their sandbox (PR #349, thanks lukeitsback).
• The installer accepts a /Library folder whose access-control list only takes access away, and names any entry that would let another account write there.
• When a kext you added to /Library/Extensions stops macOS from building its kernel collection, the installer names that kext and says how to move it, instead of only "test kernel collection build failed".
• The package carries the licence texts of everything in it (NVIDIA's firmware licence, the open kernel modules' licence and Mesa's), installed to /Library/NullMoth/licenses; the uninstaller removes them.

• NVIDIA's driver options given as nvregdwords= no longer stop the startup. From 2.5.0 to 2.5.2 any option set this way ended the boot with a kernel panic in the driver; the options now reach NVIDIA's registry, as the 2.5.0 notes describe.
• HDMI and DisplayPort audio, stated precisely: on macOS 15 the driver gives AppleHDA the display's audio link, and the display shows up under Sound > Output. macOS 26 removed AppleHDA, so on macOS 26 this package does not provide display audio by itself.

• HDMI and DisplayPort audio: a connected display shows up under Sound > Output with no setup. The driver publishes the display-to-codec link AppleHDA needs, keeps it current through hot-plug and monitor swaps, and lets a machine whose display audio is on another codec pin name it (PR #284, thanks Hi-Jiajun; PR #313, thanks Ersiter).
• Laptops whose NVIDIA bring-up takes longer than 40 seconds keep their desktop: the startup hold waits while the GPU is still coming up, up to 120 seconds (nvrmholdcap=<seconds> sets it) (PR #317, thanks watchthelight).
• Machines with several NVIDIA cards can pin the desktop to the card that drives the boot screen, and the choice survives updates (opt-in, see docs/AUTO-PRIMARY-GPU.md; PR #247, thanks is-pan).
• High-refresh displays: sudo sysctl debug.nvrmfb_flip_latch=2 checks each flip at the next present instead of waiting for it, giving back about half a refresh period per frame (opt-in; #300, thanks xiaoming-crypto).
• The system log is quiet again: the driver logs failed GPU allocations, not every successful one (#242).
• Apps using Vision barcode detection (QQ and others) and Geekbench's Background Blur no longer sit at full CPU for minutes compiling one MPS kernel (#334, #111). That kernel still does not run on the GPU yet.

• Apple TV plays video again, and dragging windows no longer kills apps like Telegram: CoreAnimation's extended-range video formats (BGRA10_XR / PBGRAXb and BGR10_XR) are supported; the driver refused them, so video and some window pipelines failed (#309, #280).
• Shader writes to sRGB textures land, encoded as on a real Mac; they were silently dropped, which broke Blender's edit-mode viewport (#321).
• Scaled and HiDPI desktop modes work: a desktop size different from the display's timings is scaled by the GPU's head scaler instead of being refused (PR #273, thanks Leaf-lsgtky and lsCoding666, #221).
• Faster window updates: CPU-drawn surfaces upload from their own pages through the copy engine (PR #296, thanks mickys).
• GPU temperature in iStat and other monitors on cards whose sensor register answers (tested on Ampere; PR #323, thanks GeekAudio).

• Built on NVIDIA's 615.78.08 driver (from 610.57.04): NVIDIA's own kernel driver, display code and GSP firmware for every card, with all of NVIDIA's fixes since 610, including the blocky corruption in The Last of Us Part II.
• New cards: NVIDIA T400, T500, T550, T600, T1000 and T1200 (desktop and laptop), and the second RTX 5070 device ID.
• The GPU's identity now comes from the PDI NVIDIA stores in the card's PCIe configuration space, the way NVIDIA's driver reads it; older boards fall back to the previous method.
• NVIDIA's driver options work on macOS: the boot-arg nvregdwords=Key=Value;Key=Value is NVIDIA's NVreg_RegistryDwords, read by NVIDIA's own parser. nvregdwords=RmDisableDisplayGlitchPerfLimit=1 is NVIDIA's new lower idle power with several monitors (it can glitch the display briefly when clocks change).
• 12-bit HDMI deep colour: the boot-arg nvmaxbpc=12 allows it on displays that report it. The default stays at 10 bits, as NVIDIA ships it.

• Hardware H.264 and HEVC encode on every GTX 16 card: GTX 1650, 1650 SUPER, 1660, 1660 SUPER and 1660 Ti use the same encoders as RTX 20.
• HEVC encode on GTX 16 and RTX 20 cards keeps going past the second frame; it stopped with an error from frame 2 on.
• Laptops whose screen is on the Intel or AMD GPU offer the NVIDIA Metal device about 8 seconds after startup instead of 60, so apps opened at login see it too.
• WindowServer no longer crashes in the Vulkan driver when a binding holds a handle that cannot be a live object: the slot gets a placeholder and the log names it. Two RTX 3060 machines restarted WindowServer a few seconds after each login on this.
• An app no longer freezes for good when one finished GPU job is missing its completion mark (#322): Terminal and iTerm2 printing fast stopped redrawing and taking input until force-quit.
• Shaders that call functions through a visible or intersection function table load again; they returned nothing since 2.3.6 (PR #318, thanks watchthelight).
• World of Warcraft starts again: since 2.3.6 it showed a black screen and "unable to start up 3D acceleration", and once past that the terrain lighting shader faulted (PR #304, thanks razvanccr).
• Apps built on wgpu (Tern and others) start again: a fragment shader that clears the padding of a local struct failed to translate and the app stopped at launch (#330).

• Fix a desktop freeze followed by a kernel panic ("thread_call group 'high' reached max thread cap") a few seconds to a few minutes after logging in (#295). The display driver started a new kernel thread for every vblank and every cursor move while an earlier one was still waiting; each of its callbacks now runs one at a time.
• Metal apps translate a shader they have already seen without hashing it again: the digest kept on the translated text was never reused (PR #294, thanks watchthelight).

• Fix the black screen and the startup that stops on the verbose text after updating to 2.4 or 2.4.1. The display driver checks for a part of the display interface that the 2.4 and 2.4.1 core driver was built without, so it refused to start and the screen never moved off the startup text. 2.4.2 builds the core driver with it again; nothing else changes.

• Fix a black screen after updating to 2.4. The 2.4 package carried a framebuffer kext built from an outdated copy of its source, missing the display code the earlier releases had; 2.4.1 ships the complete framebuffer with the 2.4 changes.

• Hardware H.264 and HEVC video encode on RTX 20 (Turing) and RTX 30 (Ampere) cards, selected automatically per card; earlier builds used the hardware encoder on RTX 50 only.
• Security hardening: caller values are bounded, each client is serialised, and kernel addresses, registers and firmware stay private.
• macOS 26: the system icons and the glass menus and panels draw correctly again.
• Install from a root-owned copy of the package, with the previous driver kept as an automatic fallback.
• The kernel driver reads the power source (battery or mains) and reports it to the GPU firmware.
• Clear a stale shader cache left by an earlier build, which could show a black screen at login after an update.
• A boot counts as good once the desktop is up.

Desktop and windows

• Dragging and scrolling windows: a surface whose video memory had been set aside is put back before its update is recorded, so the driver no longer copies the whole window back to system memory on every frame (PR #245, thanks HighDelay; #253 measured the same readback on an RTX 5070 Ti at 4K)
• Moving the mouse no longer stalls the compositor: the cursor position is applied on its own thread instead of under the system's cursor lock, and OpenGL apps wait for the display's real refresh rate instead of 60 Hz on faster panels (thanks pengubow)
• Core Animation passes that read the colour attachment with framebuffer fetch off now read a defined image instead of an unbound descriptor, which could put garbage on screen inside WindowServer (PR #270, thanks watchthelight)
• Text and badge masks drawn into A8 targets land in the right lane, and Dock badges show their numbers again (PRs #271 and #274, thanks watchthelight)

Apps and games

• OpenGL: Minecraft's vertex shaders translate, the draw path no longer allocates per draw, texel-buffer descriptors and unchanged bindings are reused, depth bias and per-instance step rates take effect, samplers work inside argument buffers (thanks HighDelay)
• Chrome and Chromium use Metal through ANGLE by default, so their GPU process starts; Electron apps stay opt-in with NVMTL_ANGLE_METAL=1, and NVMTL_ANGLE_METAL=0 turns it off everywhere (PR #263, thanks Hi-Jiajun)
• Shaders with [[visible]] function references that function constants turn off are no longer refused, which lost pipelines in the login window, Spotlight, icon services and the Dock
• Metal reports the card's whole video memory as its working set, as a Mac does (8 GB cards said 5.9 GB, 12 GB cards 8.9 GB)

Stability

• The login screen no longer stays black for a minute after the display comes up. The driver waited 60 seconds before letting the compositor's frames reach the panel, on the assumption that the screen was already showing them another way; on macOS 26 it was not, so every boot sat on a black login screen for exactly that minute

Pre-release for the early-updates testers; it becomes the general release only after every GPU generation passes the field gate.

Stability

• Window dragging, Launchpad and games stutter less: each screen update waits at most one refresh period for the previous one, instead of a fixed tenth of a second; the wait wakes on the display's own event
• macOS Sequoia screenshots and window captures: the capture ordering fix from 2.3.2 returns, and it now follows what the installed kernel driver supports, so a Metal driver and a kernel driver from different versions never mismatch again
• Turing (GTX 16 and RTX 20): the login screen no longer stays black with a cursor when the first frame has not been drawn yet; the driver only shows video memory the Metal driver has written
• Turing: the ten-second start wait the installer used to add is now the driver's own default, so a laptop card or a Mac whose boot arguments were rewritten gets it too (a value you set yourself still wins)
• Laptops with a hybrid display: the driver no longer copies the boot screen from another card's memory, which could stop the Mac during startup. When WhateverGreen is present next to the driver, the installer notes the gfxrst setting it needs
• A Mac reset during the driver's own start-up wait (the ten-second Turing settle, the 100-second wait when the card's memory window cannot be moved) no longer counts toward switching the driver off. Two such resets used to leave a working Mac on the fallback driver, reported at the next login as "driver loaded: no"
• A GPU that is busy is not a GPU that is gone: the Metal driver keeps waiting for a late frame instead of disabling itself for the rest of the process. In WindowServer that disablement meant every later command buffer was refused and the session restarted ("Failed to get a MTLCommandBuffer")
• Every kernel driver now carries the package version. A Mac that still runs an older kernel driver after an update (the kernel collection was not rebuilt) is named in Send logs and in the "reached the desktop" report instead of counting as the new version

Logging (so a report says what happened)

• The kernel driver logs its own version at start
• A full video-memory window logs the first refusals and then every 256th, with one line saying who holds the window
• Spare framebuffer instances (more instances than displays) say so instead of reporting a failure
• A display that stops flipping logs how long the frame had been pending and the flip counters
• Every ten seconds of screen activity, one line per display records the flip rate against the panel rate, missed refreshes and skipped frames
• A refused mode set or flip in the kernel names the stage that refused it
• The Metal driver says why it could not begin a command buffer (a lost GPU, a failed allocation), the line WindowServer crashes without

Stability

• 2.3.2 made some Macs worse: blank or torn windows and a Dock missing its apps on macOS Sequoia, WindowServer crashes on some Tahoe desktops. Its display changes (screenshots and the lock screen, the brightness and video-memory logging, the Turing flip change) are withdrawn from the package until they pass a staged test with the testers. 2.3.3 ships the 2.3.1 driver, which has the longest clean record in the field, on both macOS 15 and macOS 26
• macOS Sequoia: installing and updating works again (the 2.3.2 fix is kept: the Metal driver is built for macOS 15)
• After a rollback, the installed version record follows the files that are really installed (kept from 2.3.2)

Fixes

• macOS Sequoia: installing or updating to 2.3.0 or 2.3.1 stopped with "runtime payload requires macOS 26.0.0". The Metal driver in those packages was built for macOS 26 only; it is built for macOS 15 again, and a release can no longer ship anything that needs a newer macOS than 15
• Screenshots, window captures and the lock screen showed a blurred or empty window. macOS reads a window right after drawing it, and the driver now finishes the drawing first, as Apple's drivers do
• Send logs now includes how much of the card's video memory window is in use, so a black screen with only a cursor can be told apart from a leak
• After an update was rolled back, Send logs still named the new version. The installed version now follows the files that are actually installed, so reports say what your Mac is really running
• The Metal driver never hands the GPU a missing texture or sampler; it uses an empty one and logs it (a cause of Safari crashes in older versions)
• Laptops with the screen on the NVIDIA GPU: every step of the brightness setup is logged, so brightness that stays greyed out can be traced from Send logs

AMD / Intel test program

• A test driver that fails to start no longer panics the Mac: memory still mapped when the start gave up is kept until it is unmapped, instead of being freed under the mapping
• Intel 11th and 12th gen graphics (Tiger Lake, Rocket Lake, Alder Lake-S) and DG1 now reach the Intel driver. Two drivers list these chips and the device was offered to the wrong one, so neither took it
• When a start fails, its first error lines are kept and reported on the next start, even if the Mac stopped
• The test driver's log names AMD and Intel devices correctly

Safer updates

• If an update cannot bring your display up, your Mac no longer stays on a black screen. After two starts in a row that never reached the display, the driver turns itself off so macOS still starts, and the driver you had before the update is put back automatically with one restart. Nothing stays running once an update has worked
• Releases now also have to pass app checks on the new Metal driver (Terminal text, window capture) and may not do worse than the release before

Fixes

• Plugging a different monitor into the same port: macOS sees the new monitor (its name, size and modes) instead of keeping the one it replaced
• Terminal and other apps that draw text with programmable blending: text no longer comes out missing or garbled (GTX 16 and others). Draws that read colour from a pass without every attachment were skipped; they are now drawn, and overlapping triangles in one draw read each other's colour in order (thanks pengubow)
• Above 60 Hz: the display and the accelerator woke up once per millisecond to check for a finished flip, which users' flicker recordings showed as about 56 flips a second on a 143 Hz panel while windows moved. They now wake on the flip itself (the display on macOS 15 and 26, the accelerator on macOS 26)
• When a display's mode set is refused, the driver now tries the mode the firmware is already running before giving up, and the log says why the mode was refused (reports of a black screen at login on some RTX 3060 boards)
• Draws the Metal driver had to drop are now in the logs that Send logs collects

AMD / Intel test program

• 1401 can join the AMD and Intel GPU test program (AMD/Intel panel, Join). Testers get the newest signed test build of the AMD/Intel driver, and after every restart 1401 sends what the driver did, so a start that hangs or panics still reaches us. Leave the program from the same panel at any time

This release makes the driver do one thing: drive NVIDIA GPUs. Everything else it used to carry is now its own product or is gone, which makes the package smaller, leaves less running on your Mac, and leaves less to go wrong.

Leaner

• The package carries the NVIDIA driver only: the AMD/Intel GPU driver, the NPU driver and the audio driver are separate products with their own packages and installers (the package drops from about 1 GB)
• The boot logo is gone: the OpenCore logo driver, the boot progress screen and the display driver's code that kept the logo on screen. Installing, updating or removing the driver takes an existing logo out of OpenCore
• Three background jobs are gone: a display tracker that streamed the driver's log all the time, a WindowServer restart at every boot (a flicker workaround from before the flicker fix in 2.1.0), and a crash-check login item. Logs are collected only when you press Send logs. Installing or removing the driver takes out what an earlier version left
• The 1401 Mac app moved to the 1401 project; it installs this driver with the package's own installer

Fixes

• A display whose firmware boot timing NVIDIA's display code does not accept (seen on an RTX 3060 at 1920x1080) now switches to the validated timing for the same resolution instead of crashing WindowServer at login
• Checksum lists are checked strictly: a malformed line in SHA256SUMS used to be skipped while the check still passed
• install.sh fixes PATH before running anything as root, like uninstall.sh
• Crash reports sent with Send logs keep the library identifiers needed to read them (machine identifiers are still removed)

Driver

• Each GPU generation (GTX 16, RTX 20, 30, 40 and 50) is now described by the card itself: the driver reads the architecture and the newest engine classes from the GPU and publishes them, and the Metal driver picks the video decoder and the scan-out model from that instead of a table of device IDs
• Cards with a small BAR1 (Resizable BAR off, or a board that does not offer it): the driver moves BAR1 away from the boot screen at its own size, so GTX 16 / Turing cards get the same zero-copy display path as the others. A BAR1 that macOS left unassigned is reported instead of failing with no reason
• The desktop no longer waits for the GPU on every frame (#196): each window the compositor draws is marked when the GPU finishes it, and the display waits only for that mark
• A BAR1 the firmware already placed large is left as it is, and the old setup comes back if a resize fails
• Gamma tables are applied; HiDPI desktop modes can be turned on with nvfbhidpi
• The installer keeps your own settings block in nvrm610.conf

Metal

• Blender: Eevee renders lit scenes. NVIDIA GPUs cannot write the stencil value from a shader, which Blender's Metal backend relies on to sort its lighting; the driver now rebuilds that value by drawing once per stencil bit. Also fixed: a light loop that hung the GPU (the first active SIMD lane, and loop values lost when helpers were inlined), render passes with no attachments, layered 3D colour targets, framebuffer fetch, samplers held in argument buffers, and image atomics in helper functions. Blender's shadow pages no longer crash it
• Indirect compute dispatches read their grid size on the GPU instead of stopping to read it on the CPU
• Hardware video decoders are registered with macOS's decode service (#175)

AMD / Intel test driver

• The log now says which driver started, which PCI drivers registered and which device each one matched, so a GPU no driver took can be traced from a user's log

Mac app

• Map my GPU (experimental, opt-in)

Mac app

• Prepare this Mac for Tahoe is back: on macOS 15 one button installs or updates the driver (it carries the Tahoe driver too), parks it and sets the installer's GPU settings, then opens Software Update. After the update the first start sets the driver up for Tahoe and restarts once by itself. Cancel puts everything back, also with a 1401 stick plugged in next to the internal disk (it changes back the config the preparation changed)
• Intel and AMD laptops and PCs whose own GPU OpenCore switches off (disable-gpu, or -wegnoigpu on Intel) can try the AMD / Intel test driver: the app no longer says "no GPU" and stops; turning the test driver on turns that GPU back on (also on Core Ultra, where the NPU used to hide it), and turning it off switches it off again exactly as it was

1401 for Windows

• When the stick's disk number is gone at write time, the log now lists the USB disks Windows shows instead

Driver

• System sleep and wake (PR #182, thanks danlt150): the GPU's memory is kept across sleep, the PCI setup and Resizable BAR are restored on wake, and the display driver does the wake steps macOS expects. Before, the GPU lost its state in sleep and the desktop could freeze on wake
• The Metal driver finds a submission's copy buffers once instead of on every check (PR #223, thanks HighDelay)
• The display driver no longer fills the log with the same "hdcs" line many times a second (#216)
• Security (#214): diagnostic sysctls that let any user read physical addresses and memory are gone from normal builds (they exist only with the nvrmdiag=1 boot argument, and only root can read them); starting the GPU by hand is root-only; the kernel addresses used for debugging are left out of shipped builds

Mac app

• AMD boards (AM4 and AM5) whose OpenCore config rewrites the firmware memory map get DisableVariableWrite, so macOS cannot write NVRAM through that map (boards that stopped starting after an install)
• When no NVIDIA card is visible, the app says why: -wegnoegpu or -wegonogpu in boot-args, a disable-gpu property, or a laptop BIOS set to integrated graphics only
• Macs with only an AMD or Intel GPU can turn on the experimental AMD / Intel GPU driver; it installs from the downloaded driver package, without the NVIDIA driver
• Laptops whose built-in screen is wired to the NVIDIA GPU get an opt-in switch for brightness control

1401 for Windows

• Intel PCs (12th gen and newer) that stop at EB.MM.AKM because no kernel slide fit: rebuilding with the stick plugged in reads its startup log and turns on DevirtualiseMmio for the next start
• AMD PCs get the same NVRAM write guard as the Mac app (DisableVariableWrite when the memory map is rewritten)
• A stick whose disk number changes during the write is found again instead of stopping the build; a volume that disappears is skipped
• When Apple's download server resets the connection, the stop message says what to try on the network
• When Windows cannot format the stick, its own error is now in the log before the fallback

1401 for Windows

• A PC with no network adapter that works in the macOS installer gets its stick written, with a note to connect a supported Ethernet or Wi-Fi adapter before installing
• A stick that Windows gives a new disk number just before the erase is found again and written

No changes to the driver or the Mac app since 2.0.0.

1401 for Windows

• PCs with an NVIDIA card can build for macOS 26 Tahoe. 2.0.0 offered only macOS 15 for them, so picking Tahoe stopped the build

No changes to the driver or the Mac app since 2.0.0.

macOS 15 Sequoia and macOS 26 Tahoe.

Security

• Fixes for every finding in the security report by @Mahasvan (#214). Thank you for reporting them to us first.
• The driver's administrator check now reads the caller's real user id; it used to answer yes for everyone
• Copies between NVKMS and a process can no longer reach kernel memory, and the NVKMS entry point is root-only
• First-time GPU bring-up requires root
• A process cannot map memory that belongs to another process
• Unknown client types are refused, the bind call checks its range and always drops a stale mapping, waits on GPU events are capped, and one process can no longer fill the client table (slots are kept for the system)
• The kernel-address check refuses to arm on any macOS build it was not made for
• The Mac app verifies a signature on its own updates before unpacking them, runs with the hardened runtime, and refuses to run its root scripts from a folder other users can write to

Driver

• GeForce RTX 50 series: compute kernels compiled by NVIDIA's own compiler now run on the card
• Fix the black screen with only a cursor after updating to 1.15 (video memory reserved over the boot screen was handed out again), including displays larger than the boot screen's reserved memory, such as 3840x2560 on cards with a 256 MB memory window
• Fix flicker caused by a reused buffer showing an old frame
• A second monitor keeps its high refresh rates, and panels faster than 240 Hz are paced correctly
• Ultrawide, 4K and HiDPI modes are no longer capped at twice the boot resolution
• Cards with a small memory window: the display's own share is counted, so the desktop stops running out of space; a Turing card left at 64 MB is resized to the largest size it offers
• A display connected after startup comes up without a restart (#142, thanks @lsCoding666)
• macOS learns the panel's physical size, so scaling is right instead of assuming 72 dpi (#199, thanks @GeekAudio)
• Laptops with the screen on the Intel GPU: the driver no longer takes the Intel GPU's display slot (#131, #119, thanks @TararaDotJoshua), and muxless Turing GPUs get the startup settle they need
• Tahoe: ship the macOS 26 accelerator built from this release (1.15 carried an older one: black screen on Tahoe)
• Fix crashes in Photos and media analysis, in apps that split a command buffer at an event, in Chess, and in apps that free Vulkan memory (thanks @Exop63)
• Fix memory running out during Super Resolution and other long machine-learning work
• A texture copied for display stays in step on every submission (thanks @KolliasG7); work is reported as scheduled only after IOSurface writes finish (#147, thanks @imyago9); a surface keeps its video memory across page-off (#200, thanks @imbrahiam)
• World of Warcraft Classic and other Metal games: cube-map arrays, bounds-checked buffer reads and several shader fixes (#171, thanks @razvanccr); Firefox WebRender shaders translate (thanks @advokat11); MSAA sample positions on cards without programmable 1x positions (#136, thanks @eisda)
• Faster pipeline builds: built in parallel, one hash per shader module, planner results reused, and one private image per IOSurface (#173, #177, #181, #188, #189, thanks @modesttao)
• Chrome, Edge and Electron apps draw on the CPU again by default, which ends the artifacts, stutter and voice-call crashes in Discord on several cards; NVMTL_ANGLE_METAL=1 turns their GPU path on
• Optional asynchronous texture uploads, off by default (NVMTL_ASYNC_TEX_UPLOAD=1)
• Multi-GPU display selection workaround documented in docs/MULTI-GPU.md (#197, thanks @is-pan)

Experimental, off unless you turn it on

• AMD and Intel GPU support (NMXGPU) and the NPU driver, for macOS 15 and 26: Setup > "AMD / Intel GPU driver (test)"

Mac app

• Finish a macOS update the update helper never finished, and keep this Mac's own BAR settings afterwards
• "Remove NVIDIA driver" stops with an explanation, instead of failing half-way, when System Integrity Protection is on
• The Remove NVIDIA driver startup entry is always visible, and removing the driver keeps the screen on
• Warn when the Mac starts OpenCore only from the USB stick and the internal disk has none
• Explain a failed config backup on a read-only OpenCore partition, and say which step comes first when Install is greyed out
• Send logs copies files before running log queries and includes crash reports under any name

1401 for Windows

• Pick macOS 15 Sequoia or macOS 26 Tahoe before the build; see the 1401 changelog for the full list

Known issues

• Preview not showing pictures and lost window transparency on some macOS 15 setups
• Chrome decodes video on the CPU while its GPU sandbox is on (#175)
• Thunderbolt eGPUs: no display
• Laptops with the screen on the Intel GPU: the built-in screen is not accelerated
• HDMI/DisplayPort audio: the display now reports its audio formats to macOS, but sound over HDMI/DisplayPort is not working yet

The last release for macOS 15 Sequoia.

Driver

• Fix black screen with only a cursor on GTX 16 and other cards without Resizable BAR (regression in 1.14)
• Fix Unity 6 games hanging on the first frame
• Fix windows and web pages showing old content after a redraw
• Fix high WindowServer CPU and low frame rate at 4K after a few minutes
• Fix slow scrolling (window copies up to 35x faster)
• Fix monitors shown as a 30-inch TV in About This Mac
• Fix 165 Hz mode missing on some monitors (since 1.14.2)
• Retry a display that stays dark after sleep on the next wake
• Fix Chrome, Edge and Electron apps drawing everything on the CPU (no WebGL, slow scrolling); their GPU path is on by default again and NVMTL_ANGLE_METAL=0 turns it off
• Allow Chrome, Edge and Electron apps to use the GPU with their sandbox on
• Fix 5120x1440 and other ultrawide modes missing when the firmware started at a smaller resolution
• Fix green lines in Photos and other VideoToolbox scaling (#179, #180, thanks @modesttao)
• Add T400, T600, T1000 and T-series laptop GPUs (#132, thanks @TararaDotJoshua)

Mac app

• Show download progress in Update driver, which looked stuck at "Downloading the newest driver..."
• Ask for access to USB sticks instead of failing with "no OpenCore config"
• Say when the boot stick is unplugged
• Recover from an interrupted install instead of stopping at "installer guard already exists"
• Download from the NullMoth mirror when GitHub is blocked
• Send logs: always finish, keep the start of each boot, stop re-sending old logs

1401 for Windows

• Fix the build stopping on ACPI tables that declare a method twice (BIOSTAR B650 and similar)

Other

• Grammar fixes in the release notes (#128, #129, thanks @lunar-me)
• Translator test fix (#172, thanks @modesttao)
• Release notes moved into this file; add CONTRIBUTING.md and a pull request template

Known issues

• If Chrome pages show corrupted areas on your card, start it with NVMTL_ANGLE_METAL=0 and send logs
• Ghost or black areas after windows overlap with a BAR of 4 GB or more (#127)
• Chrome decodes video on the CPU while its GPU sandbox is on: VideoToolbox picks a decoder before the driver loads (#175)
• Thunderbolt eGPUs: no display
• Laptops with the screen on the Intel GPU: the built-in screen is not accelerated
• No HDMI/DisplayPort audio

• Stop driver updates from turning the EFI boot logo back on
• Keep the existing OpenCore BAR settings when installing
• Show install, cancel and failure in the updater after the download finishes

• Fix Firefox WebGL inside the GPU sandbox
• Turn off the Chromium ANGLE Metal path by default after corruption reports on RTX 30 and 50
• Retry Apple recovery downloads that return HTTP 403

• Fix Unity 6 games hanging on the first frame
• Keep Chromium WebGL on the GPU when the sandbox blocks a memory query
• Send logs: give each log query its own time limit

• Fix full-screen flicker on machines that copy frames to the screen
• Send logs includes the AMD/Intel test driver

• Stop macOS writing to firmware memory on Ryzen 7000/8000/9000 (AM5)

• Fix black login window on high refresh rate monitors
• Fix SecurityAgent and Installer crashes on some RTX 40 PCs
• Fix apps quitting on a GPU check (Safari, photo analysis)
• Fix the AMD/Intel test driver firmware download

• Fix PCs that stopped starting after 1.14.0
• Fix the boot logo and test driver switches doing nothing

• Fix black screen with a cursor about a minute after start-up (#16, #26)
• Fix 4K at 120 Hz and other DSC modes going black (#77)
• Turn the monitor off when the display sleeps (#90)
• Bring a display back after switching inputs (#110)
• Fix black screen after 100 seconds on older Intel boards without Resizable BAR

• Fix Turing cards stuck at start-up after updating
• Fix "Remove NVIDIA driver" not removing the driver, or hanging
• Crash reports no longer show crashes from before the update

• Use the NVIDIA GPU on laptops whose screen is on the Intel GPU (as a second Metal device)
• Record display state from the first second of each boot for flicker reports
• Fix Live Text crashing in the Vulkan driver

• Download Lilu and AMFIPass when they are missing from the EFI
• Fix Firefox crashing at launch (ipc_control_port_options=0)
• Fix "STOP back up nvmtl"
• Add USBToolBox when it is missing

• Fix AM5 Ryzen PCs stuck at EB.MM.AKM
• Drop the DMAR table on Intel Core Ultra
• Reach Apple's recovery server directly when DNS fails

• Try one more memory setting on AM5 boards
• Erase stubborn USB sticks with diskpart

• Use known-good settings for boards that have booted before
• Fix "Format-Volume: Invalid property" and "WinError 2" when erasing

• Pass the board's DSDT to macOS again (was dropped on about 1 in 5 PCs)
• Send logs delivers the flicker capture and OpenCore config

• Fix the driver removing itself after login on some boards
• Fix RmInitAdapter failed on Ryzen with a custom EFI
• Send logs includes the OpenCore config (serials removed)

• Finish the macOS download on networks that cut connections

• Clear flicker at login with a compositor restart
• Build on PCs whose ACPI tables crash the disassembler

• Fix Send logs freezing the app
• Fix the crash loop on 16 GB RTX 50 cards
• Light up a monitor plugged in after login

• Fix "Remove NVIDIA driver" so the driver stays off until removed
• Show the picture with macOS's display driver while ours is off

• Fix recycled allocations reaching queued GPU work
• Fail a rejected display request on its own instead of stopping the display

First public releases: installer, updater, removal and USB map fixes.