GPUs · AMD
Can the RX 7900 XT run local LLMs?
With 20 GB of VRAM (about 18.0 GB usable) and 800 GB/s of memory bandwidth, the RX 7900 XT runs 24 of 39 popular open models fully on the GPU at 8K context. The largest is Mistral Small 3 24B at Q4_K_M, at roughly 37 tokens per second.
Best picks for the RX 7900 XT
Every model on the RX 7900 XT
Results at 8K context, picking the highest quality quantization that fits. "Offload" means part of the model runs from system RAM (we assume 32 GB), which works but is much slower.
| Model | Status | Quant | Memory | Speed |
|---|---|---|---|---|
| Mistral Small 3 24B | Runs well | Q4_K_M | 18 GB | ~37 tok/s |
| Devstral Small 24B | Runs well | Q4_K_M | 17 GB | ~37 tok/s |
| Codestral 22B | Runs well | Q4_K_M | 16 GB | ~40 tok/s |
| gpt-oss 20B | Runs well | Q5_K_M | 17 GB | ~89 tok/s |
| Qwen3 14B | Runs well | Q6_K | 15 GB | ~43 tok/s |
| DeepSeek-R1 Distill 14B | Runs well | Q6_K | 15 GB | ~43 tok/s |
| Phi-4 14B | Runs well | Q6_K | 15 GB | ~43 tok/s |
| Gemma 3 12B | Runs well | Q8_0 | 17 GB | ~40 tok/s |
| Mistral Nemo 12B | Runs well | Q8_0 | 16 GB | ~40 tok/s |
| Gemma 4 12B | Runs well | Q8_0 | 17 GB | ~41 tok/s |
| Qwen3.5 9B | Runs well | Q8_0 | 12 GB | ~52 tok/s |
| Qwen3 8B | Runs well | Q8_0 | 11 GB | ~60 tok/s |
| Llama 3.1 8B | Runs well | Q8_0 | 10 GB | ~61 tok/s |
| DeepSeek-R1 Distill 8B | Runs well | Q8_0 | 10 GB | ~61 tok/s |
| Qwen2.5-Coder 7B | Runs well | Q8_0 | 9.8 GB | ~65 tok/s |
| Mistral 7B v0.3 | Runs well | Q8_0 | 9.4 GB | ~68 tok/s |
| Qwen3.5 4B | Runs well | Q8_0 | 6.4 GB | ~100 tok/s |
| Gemma 3 4B | Runs well | Q8_0 | 6.9 GB | ~110 tok/s |
| Qwen3 4B | Runs well | Q8_0 | 5.8 GB | ~120 tok/s |
| Phi-4 Mini 3.8B | Runs well | Q8_0 | 5.2 GB | ~130 tok/s |
| Llama 3.2 3B | Runs well | Q8_0 | 4.6 GB | ~150 tok/s |
| Llama 3.2 1B | Runs well | Q8_0 | 1.6 GB | ~410 tok/s |
| Gemma 3 1B | Runs well | Q8_0 | 1.4 GB | ~490 tok/s |
| Qwen3 0.6B | Runs well | Q8_0 | 0.9 GB | ~820 tok/s |
| Gemma 4 26B-A4B (MoE) | Offload | Q4_K_M | 19 GB | ~14 tok/s |
| Qwen3.6 27B | Offload | Q4_K_M | 19 GB | ~4.7 tok/s |
| Qwen3 30B-A3B (MoE) | Offload | Q4_K_M | 21 GB | ~15 tok/s |
| Qwen3-Coder 30B-A3B | Offload | Q4_K_M | 21 GB | ~15 tok/s |
| Gemma 3 27B | Offload | Q4_K_M | 22 GB | ~4.3 tok/s |
| Qwen3 32B | Offload | Q4_K_M | 23 GB | ~3.5 tok/s |
| DeepSeek-R1 Distill 32B | Offload | Q4_K_M | 23 GB | ~3.5 tok/s |
| Qwen2.5-Coder 32B | Offload | Q4_K_M | 23 GB | ~3.5 tok/s |
| Qwen3.6 35B-A3B (MoE) | Offload | Q4_K_M | 23 GB | ~16 tok/s |
| Gemma 4 31B | Offload | Q4_K_M | 25 GB | ~3.6 tok/s |
| Llama 3.3 70B | Won’t fit | — | 48 GB | — |
| DeepSeek-R1 Distill 70B | Won’t fit | — | 48 GB | — |
| Llama 4 Scout 109B-A17B | Won’t fit | — | 73 GB | — |
| gpt-oss 120B | Won’t fit | — | 77 GB | — |
| Qwen3 235B-A22B (MoE) | Won’t fit | — | 155 GB | — |
What this means in practice
The RX 7900 XT is comfortable with models up to about 20B to 27B parameters. Gemma 4 26B-A4B (MoE) is the most useful step up beyond that: it needs about 19 GB, so part of it runs from system RAM, but as a mixture-of-experts model it stays reasonably quick at around 14 tokens per second. Models from Llama 3.3 70B upward need 48 GB or more and are out of reach. AMD cards work with Ollama and LM Studio through ROCm or Vulkan; setup is smoother on Linux and recent Windows drivers.
The next memory tier up is 24 GB: RTX 4090, RX 7900 XTX, RTX 3090 / 3090 Ti.
Compare the RX 7900 XT
Questions
What is the biggest model the RX 7900 XT can run?
At 8K context the largest model that fits entirely in its 20 GB of VRAM is Mistral Small 3 24B at Q4_K_M, at roughly 37 tokens per second.
Can the RX 7900 XT run 70B models like Llama 3.3 70B?
No. It needs about 48 GB at Q4, which is far beyond 20 GB even with system RAM offload.
What is the best model for coding on the RX 7900 XT?
Devstral Small 24B at Q4_K_M fits fully and runs at about 37 tokens per second. Pair it with an editor extension like Continue pointed at Ollama.
Estimates based on memory size and bandwidth; real speed varies with drivers, software and context length. See how we calculate VRAM.