Can the Radeon RX 9060 XT 16 GB run Gemma 4 26B A4B?
Not fully, but it's usable. Gemma 4 26B A4B needs about 16.8 GB at Q4_K_M and the Radeon RX 9060 XT 16 GB has 16 GB - yet as a mixture-of-experts model it reads only ~4B parameters per token, so you can keep attention on the GPU and the experts in system RAM (llama.cpp --n-cpu-moe) and still get interactive speeds with 32 GB+ of RAM.
Every quantization
Gemma 4 26B A4B (26.5B, 4B active) on 16 GB
| Quant | Weights | Needs | Fit | Est. speed |
|---|---|---|---|---|
| Q4_K_M | 15 GB | 16.8 GB | Too big | - |
| Q5_K_M | 19 GB | 20.8 GB | Too big | - |
| Q8_0 | 28 GB | 29.8 GB | Too big | - |
| FP16 | 53 GB | 54.8 GB | Too big | - |
Estimates. Needs = weights + 1.8 GB for KV cache and runtime at modest context. Speed is single-stream decode scaled from the Radeon RX 9060 XT 16 GB's 322.3 GB/s bandwidth (Estimated from memory bandwidth (llama.cpp decode)). Long contexts need more memory and run slower.
Run it with expert offload
Attention on the GPU, experts in system RAM
llama-server -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf -ngl 99 --n-cpu-moe 99Lower --n-cpu-moe step by step until the Radeon RX 9060 XT 16 GB's 16 GB is nearly full - each expert layer moved back to the GPU speeds things up. Needs about 8 GB+ of free system RAM.
What to do instead
Two ways forward
Cheapest cards that run it
- NVIDIA Tesla P40 - 24 GB, ~$159
- AMD Radeon RX 7900 XT - 20 GB, ~$699
- NVIDIA RTX 3090 - 24 GB, ~$750
- AMD Radeon RX 7900 XTX - 24 GB, ~$849
Models that fit the Radeon RX 9060 XT 16 GB
- Qwen2.5-Coder-14B-Instruct - 14.8B, 9 GB at Q4
- Qwen3 14B - 14B, 8 GB at Q4
- Phi-4 - 14B, 8.5 GB at Q4
- Gemma 4 12B - 12B, 7 GB at Q4
Compare
Same model, other hardware - and other models on this card
Gemma 4 26B A4B on other hardware
- RTX 5090~303 tok/s
- RTX 5090 D~302 tok/s
- RTX 6000D~265 tok/s
- RTX 4090~184 tok/s
- Radeon RX 7900 XTX~162 tok/s
- RTX 3090~161 tok/s
- MacBook Pro 16" M4 Max~92 tok/s
- Arc Pro B60~77 tok/s
Also runs on the Radeon RX 9060 XT 16 GB
- Qwen2.5-Coder-14B-Instruct~21 tok/s
- Qwen3 14B~23 tok/s
- Phi-4~22 tok/s
- Gemma 4 12B~26 tok/s
- Qwen3.5 9B~31 tok/s
- Llama 3.1 8B~34 tok/s
- Mistral-7B-Instruct-v0.2~46 tok/s
- Qwen3.5 4B~62 tok/s
Frequently asked
Quick answers
Can the Radeon RX 9060 XT 16 GB run Gemma 4 26B A4B?
Not fully, but it's usable. Gemma 4 26B A4B needs about 16.8 GB at Q4_K_M and the Radeon RX 9060 XT 16 GB has 16 GB - yet as a mixture-of-experts model it reads only ~4B parameters per token, so you can keep attention on the GPU and the experts in system RAM (llama.cpp --n-cpu-moe) and still get interactive speeds with 32 GB+ of RAM.
How much VRAM does Gemma 4 26B A4B need?
About 15 GB of weights at Q4_K_M, 28 GB at Q8_0 and 53 GB at FP16. Add roughly 1.8 GB for the KV cache and runtime at modest context lengths.
How fast is Gemma 4 26B A4B on the Radeon RX 9060 XT 16 GB?
It doesn't fit in the Radeon RX 9060 XT 16 GB's memory, so speed depends on how many layers you offload to system RAM - typically single-digit tokens per second or lower.