For years the answer to "can I run an NVIDIA GPU on a Mac?" has been a flat no. Apple dropped NVIDIA support back around the Mojave era, the Hackintosh community worked around it with increasingly fragile patches, and the conventional wisdom hardened into an eulogy: NVIDIA on macOS is dead, move on.
Last week that changed. A team called NullMoth shipped a Metal driver for NVIDIA GeForce cards on macOS 15 Sequoia — and it's not a hacky proof-of-concept. It's a real driver, open source, with an installer app, a shader translator, and a Vulkan back end. It's already past 2,900 stars in its first week, and it's the single most interesting thing to happen to Mac GPU compute since Apple Silicon shipped.
But here's the part every breathless headline is glossing over, and it matters enormously for the local-LLM crowd: this driver is x86_64 only. It runs on Intel Macs and OpenCore Hackintoshes — not on Apple Silicon. So before you go shopping for an RTX 4090 to strap to your M-series Mac mini, read the caveats. This is the honest assessment I'd have wanted before I got excited.
What Actually Shipped
The repo is nullmoth/nvidia-macos-driver, and it's a full driver stack, not a stub. The description is one sentence — "Metal driver for NVIDIA GeForce cards on macOS Sequoia" — and the README backs it up with a real architecture diagram:
Metal app ─► NVMTLDriver.bundle ─► translator (Apple AIR → SPIR-V) ─► NVK (Mesa Vulkan + NAK compiler) ─► kexts ─► GPU
In plain English: it plugs into Metal the way Apple's own drivers do, then translates Metal's shader intermediate representation into SPIR-V, hands it to Mesa's NVK Vulkan driver (with NVIDIA's open-source NAK shader compiler), and finally drives the card through kernel extensions. The kernel side runs NVIDIA's own open GPU kernel modules (the r610 series) under macOS.
What that nets you, per the README:
- Metal 3 — argument buffers tier 2, ray tracing, mesh shaders, Metal Performance Shaders, and the MetalFX upscaling path
- OpenGL through Apple's GL-on-Metal layer
- OpenCL
- Core Image and Core ML
That last one is the sleeper. Core ML on a real NVIDIA card means Apple's on-device inference stack — the thing that powers everything from local image classification to the Neural Engine-adjacent work — now has a discrete-GPU target with a lot more VRAM and raw compute than any Intel-era Mac shipped with.
The Cards That Work (and the Fine Print)
Support is Turing and newer: GTX 16, RTX 20/30/40/50, TITAN RTX, Quadro RTX, and the T-series. The team developed on an RTX 5060, and RTX 5070/5080 are reported working. Pascal, Maxwell, and Kepler are explicitly out of scope — they need a different kernel/firmware implementation.
The honest caveats, straight from the README:
- Intel Macs and OpenCore PCs only. The driver targets macOS 15 Sequoia on x86_64. Apple Silicon is not on the roadmap for this — the whole point is giving Intel machines a path to discrete NVIDIA compute.
- You need OpenCore, SIP off, and a reboot. This is a kernel-extension driver. The
csr-active-configvalue disables unsigned-kext enforcement, and there's a pile of boot-args and quirks to set. The1401installer app handles most of it, but this is not a "drag to Applications" install. - Above 4G Decoding ON, CSM OFF in BIOS, plus a compatible SMBIOS model (
iMacPro1,1is the tested one). - Laptops are a maybe. A card being in the supported-device table doesn't mean the display wiring and firmware BAR setup will cooperate. Optimus-style display muxes are not magically implemented.
- The license is "Other." It's open source, but not OSI-approved — the shader translator is LGPL-3.0, the kernel modules are NVIDIA's own, and the overall repo uses a non-standard license tag. Worth a skim before you build a product on it.
None of this makes it a toy. It makes it exactly the kind of thing that lives in the Hackintosh and Intel-Mac homelab niche — powerful, real, and unapologetically not for everyone.
Why This Matters for Local LLMs Specifically
Here's where it gets interesting for the audience that actually cares — the people running big models on their own hardware.
My whole local-LLM cluster has been built around a hard ceiling I've written about repeatedly: on Apple Silicon, your VRAM is unified memory, and there's no NVIDIA or eGPU path to break past it. That's been true since the M1. It's why the "run a 125B model at home" conversation always turned into "buy an RTX 4090 and put Linux on a PC."
This driver doesn't change the Apple Silicon story — I have to be blunt about that, because I run an M4 Mac mini and the headline got my hopes up too. An M-series Mac cannot use this. The unified-memory architecture has no PCIe NVIDIA slot to drive, and the driver is x86_64 only.
What it does change is the other half of the Mac homelab:
- Intel Macs with Thunderbolt eGPUs. An old iMac or Mac Pro — or a Thunderbolt 3 eGPU enclosure — now has a real Metal target again, with NVIDIA compute behind it. An RTX 4060 or 4080 in an enclosure is suddenly a legitimate local-inference box running macOS.
- Hackintoshes. The OpenCore community has wanted a clean NVIDIA path for a decade. This is it. A well-specced Hackintosh with an RTX 50-series card now runs macOS and gives you CUDA-class hardware via Metal/Core ML.
- The "one machine for everything" dream. macOS for the daily driver, NVIDIA for the inference, no dual-boot into Linux required. For a lot of homelabbers that's the whole pitch.
The VRAM math is the headline: an RTX 5090 has 32 GB, an RTX 4090 has 24 GB, and even a 4060 Ti has 16 GB — all of it now addressable from a Mac that can also run Ollama's Metal backend, Core ML, and the rest of the macOS toolchain. That's the ceiling-breaker the local-LLM crowd has been waiting for, just not on the hardware most of us assumed.
The Hardware, If You Want to Try It
If you've got an Intel Mac, a Thunderbolt 3 eGPU enclosure, or an OpenCore box and you want to test this yourself, the cards I'd actually buy for a budget build are the same ones I recommend across my local-LLM guides:
- NVIDIA RTX 4060 Ti 16GB — 16GB of VRAM at the best price-per-GB in the current lineup, and more than enough to make a Thunderbolt enclosure a genuinely useful local-inference box. Check price
- NVIDIA RTX 3060 12GB — if you're building on a used or older Intel Mac, this is the value play; 12GB handles 13–14B models fully offloaded and it's still findable around $200 used. Check price
(Heads up: this driver targets Intel Macs and Hackintoshes — an Apple Silicon M-series machine can't use it, as covered above. And you're signing up for OpenCore, SIP off, and unsigned kexts, so don't do this on your only daily-driver machine.)
The Honest Verdict
I'm not going to tell you to flash OpenCore and disable SIP tonight. The driver is a week old, developed primarily on a single card (RTX 5060), and the "Other" license plus unsigned-kext requirement means you're trading a big slice of Apple's security model for the capability. If your Mac is your only machine and it's on Apple Silicon, this is a curiosity, not a purchase decision.
But if you're running an Intel Mac that's aging out of macOS support, or you've already got an OpenCore box, this is genuinely the most promising NVIDIA-on-macOS news in close to a decade. The local-LLM ceiling on Mac hardware just got a real hole punched in it — the kind of hole that makes a 24 GB RTX 4090 and a $200 Thunderbolt enclosure a more interesting answer to "how do I run big models at home?" than it was last month.
The caveat I'll keep repeating, because it's the one that matters most: this unlocks Intel Macs and Hackintoshes, not Apple Silicon. If someone tells you it means your M4 can run CUDA, they didn't read the README. Everything else — the Metal 3 features, the Core ML support, the shader translation — is real and impressive, and worth watching closely as the card-support list fills in beyond Blackwell.
For the first time in a long time, the answer to "NVIDIA on a Mac?" is no longer a flat no. It's "depends — but now there's a real yes for some of you."