Short answer
- Already have an RX 7900 XTX (24 GB) or RX 9070 XT (16 GB)? Use it. It runs the same models as an NVIDIA card with the same memory.
- Start with the Vulkan version of llama.cpp or LM Studio. It needs no special drivers.
- Buying new just for AI? NVIDIA is still less work, because more tools assume it.
Two ways to run: Vulkan and ROCm
Vulkan is a graphics standard every Radeon driver already has. llama.cpp and LM Studio can use it out of the box, on Windows and Linux. It is usually the fastest way to get started, and it is often as fast as anything else when the model is writing its answer.
ROCm is AMD’s own computing platform, like NVIDIA’s CUDA. It can be faster at reading long prompts, and some tools (like vLLM) need it. It is more work to install, and support differs by card and system. The current version is ROCm 10.1 (released 5 October 2026, checked 13 October 2026).
What memory gets you
Memory decides which models fit, not the brand. A 24 GB Radeon fits the same models as a 24 GB GeForce. Speed follows memory bandwidth: the RX 7900 XTX has 960 GB/s, close to an RTX 4090’s 1,008 GB/s. See what fits with the checker set to an RX 7900 XTX.
Old problems you may read about
Early in 2026, llama.cpp had bugs with newer Qwen models on ROCm. We checked them on GitHub on 10 October 2026: issue #19880 (“ROCm support for newer Qwen models broken”) was closed as fixed on 28 February, and #20024 (a memory fault with Qwen3.5-35B-A3B) on 13 March. One report, #20545 (a hang under WSL2 on Windows), was closed on 2 June without a fix. If an old forum post says Qwen models crash on AMD, update your app first — and prefer native Linux or Windows over WSL2.
Watch out for
- Ollama uses ROCm on supported Radeon cards and turns on Vulkan by default on Windows and Linux for other cards. If it still does not see your card, try LM Studio with Vulkan.
- Some newer tools only ship NVIDIA builds at first.
- Older cards (RX 6000 series and earlier) work with Vulkan, but have less memory and bandwidth.