Radeon RX 7900 XTX for Local AI in 2026: 24GB, ROCm and What Fits

The Radeon RX 7900 XTX is AMD’s 24GB gaming flagship from December 2022, and it is still the Radeon most often considered for running AI at home. 24GB is the capacity where most open models become usable at full quality, and it is the same amount of memory as an RTX 4090 or RTX 3090.

This guide covers what fits in that 24GB, how fast it runs compared with NVIDIA’s 24GB cards, what AMD software support looks like in October 2026, and when it is worth buying. BFWR has not run this card. The fit counts and estimates come from our own calculations, labelled as such; the measured speeds come from llama.cpp’s public benchmark threads, credited below.

Quick answer

• What fits: 23 of the 30 models in our database at 4-bit, the same as any 24GB card. That includes 27B to 35B models such as Qwen3.8 27B, Gemma 4 31B and Qwen3-Coder 30B-A3B.
• Writing speed: close to an RTX 4090. In llama.cpp’s public benchmarks it generates text at 167 to 191 tokens per second on a standard 7B test, against 186 to 189 for the 4090.
• Prompt speed: much slower. The same benchmarks show the 4090 processing prompts about three times faster, which matters for long documents and coding agents.
• Software: Ollama, llama.cpp and ComfyUI all support it on Linux and Windows, through AMD’s ROCm or through Vulkan.
• Price: it launched at US$999. On 17 October 2026 one price tracker had it at $1,820.99 new and $1,000 used.

The specifications that matter

All figures are from AMD’s product page. For running language models, memory capacity decides what fits and memory bandwidth decides how fast it writes; our guide to tokens per second explains why. At 960 GB/s the 7900 XTX sits between the RTX 3090 (936 GB/s) and the RTX 4090 (1,008 GB/s). The gfx1100 target is worth remembering: it is the name AMD’s and llama.cpp’s documentation use for this card.

What fits in 24GB

Our hardware data counts 23 of 30 models as running fully on the card at 4-bit. Every 24GB card gets the same count, because what fits depends on capacity, not on the brand; the RTX 3090 and 4090 score 23 as well, and so does the 32GB RTX 5090 on the same list. The extra 8GB on 16GB cards is the real step: 16GB cards such as the RX 9070 XT fit 14 of 30.

Speeds are our estimates from the formula behind our hardware pages: bandwidth × 0.65 ÷ the weights read per token. They are ceilings for short prompts, not measurements. For mixture-of-experts models, where only 3 to 4 billion parameters are active per token, that formula gives figures above 250 tokens per second, which other overheads will stop you reaching, so we do not quote them. Expect those models to be much faster than a dense model of the same size, not to hit the formula.

What does not fit: 70B dense models such as Llama 3.3 70B need about 40GB at 4-bit, so on a single 7900 XTX they spill into system memory and drop to an estimated 1 to 2 tokens per second. The full device page, RX 7900 XTX for local AI, lists all 30 models, and our 24GB VRAM page covers the tier. Leave room for context, too: a long context window adds gigabytes on top of the weights, which our VRAM calculator works out per model.

How fast it is, measured by others

llama.cpp’s maintainers run three public benchmark threads, one each for NVIDIA’s CUDA, AMD’s ROCm and the cross-vendor Vulkan backend. All three use the same test: Llama 2 7B at Q4_0, run with llama-bench. The scoreboards are maintained by the threads’ organisers from community submissions, so the results come from different machines and drivers. Here are the 24GB cards:

Two things stand out.

Writing speed is a near-tie. Generating tokens is limited by memory bandwidth, and the three cards have similar bandwidth. The 7900 XTX’s best result, 190.9 tokens per second on Vulkan, matches the RTX 4090’s. Our formula estimates about 163 tokens per second for the 7900 XTX on this 3.8GB model, so the measured results sit at or above it.

Reading speed is not. Prompt processing is limited by compute, and here the RTX 4090 is about three times faster (11,993 against 3,727 tokens per second without flash attention) and the RTX 3090 about 40% faster. In practice, a short chat feels the same on all three. A coding agent that sends 30,000 tokens of code with each request, or a document you paste in for summarising, makes you wait three times longer before the first word appears on the AMD card.

One more detail from the same data: on this card, the Vulkan backend was faster than AMD’s own ROCm in every column. The gap is small, but it means you do not need ROCm to get good speed from llama.cpp.

Software support in October 2026

AMD support used to be the reason not to buy a Radeon card for AI. It is much less of one now, and the 7900 XTX is the best-supported consumer Radeon.

• Ollama lists the 7900 XTX as supported through ROCm on both Linux and Windows. Its documentation says it “requires the AMD ROCm v7 driver on Linux” and “an AMD ROCm v7 / HIP7-capable driver stack on Windows”. Notably, the Windows list covers the RX 7000 series but not the newer RX 9000 cards, which Ollama supports through ROCm on Linux only. Ollama can also use Vulkan, which “is enabled by default when the backend is installed”.
• llama.cpp publishes ready-made Windows and Linux builds for both Vulkan and ROCm with every release. Its build guide uses the 7900 XTX as its Windows example, noting that gfx1100 “corresponds to Radeon RX 7900XTX/XT/GRE”.
• ComfyUI offers a Windows portable download for AMD cards, and its install guide lists the 7900 XT and XTX under the device-gfx1100 package for AMD’s Windows PyTorch builds. On Linux it installs PyTorch’s ROCm build with pip.
• ROCm itself: AMD’s current Linux documentation marks the 7900 XTX as supported, on Ubuntu 24.04.4 and 22.04.5 and on RHEL 10.1 and 9.7. Other distributions may work but are not on AMD’s list.

The simplest start is Ollama:

ollama run qwen3.8:27b

Then run ollama ps in another terminal; the PROCESSOR column should read 100% GPU. If it shows a CPU share, the model plus its context does not fit, and a smaller model or shorter context will fix it. For a server setup, see our guide to running Ollama as a server.

Image generation

24GB is a comfortable size for most current image models. By our calculator’s estimates, at 1024×1024 with 8-bit weights and the text encoder offloaded to system memory:

Everything except FLUX.2 [dev] fits at 1 megapixel, and FLUX.2 [dev] needs 4-bit weights to get near 24GB. These are memory estimates only; they say nothing about speed. One hardware difference is relevant here: AMD’s specification page lists 8-bit integer and 4-bit integer matrix throughput for the 7900 XTX but no 8-bit floating-point (FP8) figure, while it lists FP8 for the newer Radeon AI PRO R9700. Many 2026 image models ship FP8 files. They can still save memory on the 7900 XTX, but AMD lists no FP8 matrix rate for this card, so do not expect the speed-up those files give on newer hardware. Our guides to running Stable Diffusion locally and Krea 2 cover the setups.

Should you buy one?

The case for the 7900 XTX was always 24GB for less than NVIDIA charged. The October 2026 market has blurred that. The price tracker videocardprices.com recorded the card on one Amazon listing at $1,049.99 in June, a high of $1,899.99 on 22 September, and $1,820.99 on 17 October, with a used listing at $1,000. Prices are moving by hundreds of dollars in a week, so check before you buy.

• Buy it if you mostly chat with or generate text from 20B to 35B models, run Linux or are happy with Vulkan on Windows, and can get it near or below its US$999 launch price. Used cards are the obvious route.
• Look at NVIDIA instead if you use coding agents or summarise long documents, where prompt speed dominates; or if you rely on tools that only support CUDA, which is still common in training and research code.
• Consider the Radeon AI PRO R9700 if you need more than 24GB. It has 32GB, FP8 support and 640 GB/s, went on retail sale at US$1,299 in October 2025, and is on Ollama’s Linux supported list. Its lower bandwidth means slower generation than the 7900 XTX for models that fit both.

Our guide to the best GPU for AI covers the full range by memory tier, and the RX 7900 XTX vs RTX 3090 comparison puts the two most common used 24GB cards side by side.

FAQ

Is the RX 7900 XTX good for local LLMs?

Yes, for its capacity. It runs every model up to about 35B at 4-bit fully on the card, and generates text about as fast as an RTX 4090. Its weakness is prompt processing, which is about a third of the 4090’s speed in llama.cpp’s benchmarks.

Do I need ROCm?

No. llama.cpp and Ollama can both use Vulkan, which comes with AMD’s normal graphics driver on Windows, and in llama.cpp’s benchmarks Vulkan was slightly faster on this card. ROCm matters more for PyTorch-based tools such as ComfyUI.

Does it work on Windows?

Yes. Ollama, llama.cpp and ComfyUI all document Windows support for it.

Can two 7900 XTX cards run a 70B model?

48GB is enough for a 70B model at 4-bit (about 40GB of weights), and llama.cpp can split a model across GPUs. We have not found a published measurement for two 7900 XTX cards to quote, so we cannot say how fast it is.

Related guides and tools

• Why two 16GB cards can be 3.3x apart
• Best budget GPUs for local LLMs
• AI GPU comparison
• NVIDIA RTX 5080 vs RTX 4090 for Local LLMs
• How to Increase VRAM in 2026

Sources

• AMD: Radeon RX 7900 XTX specifications and Radeon AI PRO R9700 specifications
• llama.cpp benchmark threads: Vulkan, ROCm and CUDA (scoreboards read 18 October 2026)
• Ollama GPU documentation and llama.cpp build guide
• ComfyUI README (AMD installation) and ROCm system requirements
• videocardprices.com RX 7900 XTX price history (read 18 October 2026)
• Club386 on the R9700’s US$1,299 retail launch

Updated 18 October 2026. Prices change daily; fit counts and estimates come from BFWR’s hardware data, and measured speeds from llama.cpp’s public benchmark threads.