The Radeon RX 7900 XT has an unusual 20 GB of GDDR6 with 800 GB/s of bandwidth. That is enough for Mistral Small 3.2 24B at Q4, which 16 GB cards cannot run.
Check your own setup with the Can My GPU Run This AI Model? calculator. The numbers below come from the same data.
RX 7900 XT specs that matter for local AI
- Memory: 20 GB GDDR6 on a 320-bit bus
- Memory bandwidth: 800 GB/s
- Board power: 315 W
VRAM decides which models fit. Bandwidth decides how fast they generate text, because every new token means reading the whole model from memory.
Which AI models fit on the RX 7900 XT?
The table assumes Q4_K_M quantization (about 4.9 bits per weight) and an 8K context. Totals include the KV cache and about 1 GB of runtime overhead. Speed is an estimate: memory bandwidth divided by model size, at 60 to 80 percent efficiency, which matches published llama.cpp results on several RTX cards.
| Model | Weights | KV cache (8K) | Total (est.) | On 20 GB | Speed (est.) |
|---|---|---|---|---|---|
| Qwen3 4B | 2.5 GB | 1.2 GB | 4.7 GB | Fits | 195 to 260 tok/s |
| Llama 3.1 8B Instruct | 4.9 GB | 1.1 GB | 7.0 GB | Fits | 98 to 130 tok/s |
| Qwen3 8B | 5.0 GB | 1.2 GB | 7.2 GB | Fits | 96 to 128 tok/s |
| Gemma 3 12B Instruct | 7.5 GB | 0.9 GB | 9.3 GB | Fits | 64 to 86 tok/s |
| Qwen3 14B | 9.0 GB | 1.3 GB | 11.4 GB | Fits | 53 to 71 tok/s |
| OpenAI gpt-oss-20b | 12.8 GB | 0.2 GB | 14.0 GB | Fits | Varies (MoE) |
| Mistral Small 3.2 24B | 14.7 GB | 1.3 GB | 17.0 GB | Fits | 33 to 44 tok/s |
| Qwen3.6 27B | 17.0 GB | 0.5 GB | 18.6 GB | Tight | 28 to 38 tok/s |
| Gemma 3 27B Instruct | 16.8 GB | 1.1 GB | 18.9 GB | Tight | 29 to 38 tok/s |
| Qwen3 32B | 20.1 GB | 2.1 GB | 23.2 GB | No | – |
| Qwen3.6 35B-A3B | 22.0 GB | 0.2 GB | 23.2 GB | No | – |
“Fits” means the estimate is at most 92 percent of VRAM. “Tight” means it is between 92 and 100 percent, so close other GPU apps. Mixture-of-Experts models (gpt-oss-20b and Qwen3.6 35B-A3B) need their full weights in memory, but they only use part of them per token, so speed is not estimated.
Best models for the RX 7900 XT
- Mistral Small 3.2 24B (Q4_K_M): about 17 GB at 8K context.
- Qwen3 14B and gpt-oss-20b: comfortable, with Q8_0 of a 14B model also possible.
- Gemma 3 27B and Qwen3.6 27B: about 19 GB at Q4, so the fit is tight.
How context length changes things
The KV cache grows with every token of context, so a model that fits at 4K can fail at 32K.
| Model (Q4_K_M) | Context | KV cache | Total (est.) | On 20 GB |
|---|---|---|---|---|
| Llama 3.1 8B Instruct | 4K | 0.54 GB | 6.5 GB | Fits |
| Llama 3.1 8B Instruct | 8K | 1.07 GB | 7.0 GB | Fits |
| Llama 3.1 8B Instruct | 16K | 2.15 GB | 8.1 GB | Fits |
| Llama 3.1 8B Instruct | 32K | 4.29 GB | 10.2 GB | Fits |
| Qwen3 14B | 4K | 0.67 GB | 10.7 GB | Fits |
| Qwen3 14B | 8K | 1.34 GB | 11.4 GB | Fits |
| Qwen3 14B | 16K | 2.68 GB | 12.7 GB | Fits |
| Qwen3 14B | 32K | 5.37 GB | 15.4 GB | Fits |
A Q8 KV cache (an option in Ollama and llama.cpp) roughly halves the cache cost if you need more context.
What does not fit
Qwen3 32B needs about 23 GB at Q4 and does not fit.
Extra memory over 16 GB cards. If you want 24B models without paying for a 24 GB card, 20 GB is a useful middle. Be ready for AMD software setup.
Verdict
What speed to expect
At 800 GB/s, an 8B Q4 model has a ceiling near 163 tokens per second. Applying the 60 to 80 percent efficiency range seen in published llama.cpp tests gives about 98 to 130 tokens per second for Llama 3.1 8B and 64 to 86 for Gemma 3 12B. We did not find a reliable measured result for this exact card, so these are estimates that could be off by 10 to 20 percent. Speed on AMD also depends on the backend (HIP or Vulkan), so treat the top of the range as optimistic.
Frequently asked questions
Can the RX 7900 XT run Mistral Small 24B?
Yes, about 17 GB at Q4 with 8K context.
Can it run Qwen3 32B?
No. It needs about 23 GB.
Does it work with llama.cpp?
Yes, through the HIP and Vulkan backends.
Sources and method
- RX 7900 XT specs: Tom’s Hardware, AMD RDNA 3 / RX 7000 everything we know (20 GB, 800 GB/s, 315 W).
- Linear-attention and sliding-window layers are handled with reduced KV cache; small fixed states in some newer models are not counted.
- Model architecture: each model’s published
config.jsonon Hugging Face (Llama 3.1 8B and Gemma 3 are read from the unsloth mirrors because the original repositories require a login). - VRAM and speed figures are estimates from our calculator, not our own benchmarks. Last updated 3 Nov 2026.
