Can the Radeon RX 7900 XTX run gemma-4-26B-A4B-it?
gemma-4-26B-A4B-it·25.8B (4B active)·Radeon RX 7900 XTX·24 GB VRAM
Yes. The Radeon RX 7900 XTX runs gemma-4-26B-A4B-it fully in VRAM at Q5_K_M (about 20 GB of 24 GB), at roughly 186 tokens/s.
- best quant
- Q5_K_M
- needs
- 20 GB
- usable
- 24 GB
- est. speed
- ~186 tok/s
Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the gemma-4-26B-A4B-it page →
Every quantisation of gemma-4-26B-A4B-it on the Radeon RX 7900 XTX
| Quant | Memory | Fit | Speed |
|---|---|---|---|
| Q2_K | 11 GB | fits | ~402 tok/s |
| Q3_K_M | 15 GB | fits | ~270 tok/s |
| Q4_K_M | 18 GB | fits | ~218 tok/s |
| Q5_K_M pick | 20 GB | fits | ~186 tok/s |
| Q6_K | 23 GB | tight | ~160 tok/s |
| Q8_0 | 29 GB | offload | ~36 tok/s |
| FP16 / BF16 | 52 GB | too big | — |
Speed is a bandwidth-based estimate for generating text (960 GB/s); prompt processing and long contexts are slower.
Measured speed on the Radeon RX 7900 XTX
No community reports for this pair yet.
Report your own speed →Run gemma-4-26B-A4B-it on the Radeon RX 7900 XTX
Using Q5_K_M, the best fit for Radeon RX 7900 XTX.
ollama run gemma4:26bThe library build picks its own default quantisation.
ollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_Mdocker run -d --name ollama --device /dev/kfd --device /dev/dri \
-v ollama:/root/.ollama -p 11434:11434 ollama/ollama:rocm
docker exec -it ollama ollama run gemma4:26bThe image bundles the ROCm runtime; the host only needs the amdgpu kernel driver.
llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_MNeeds a HIP (ROCm) build of llama.cpp. Downloads the GGUF on first run.
docker run -d --name llama --device /dev/kfd --device /dev/dri -p 8080:8080 \
-v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
ghcr.io/ggml-org/llama.cpp:server-rocm \
-hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M --host 0.0.0.0 --port 8080 -ngl 99Server and web UI on port 8080.
lms get unsloth/gemma-4-26B-A4B-it-GGUF@UD-Q5_K_MThen load it from the LM Studio app or with `lms load`.
vllm serve google/gemma-4-26B-A4B-itNeeds vLLM built for ROCm; it supports Instinct MI200/MI300 and Radeon RX 7900 / 9000 cards.
vllm serve cyankiwi/gemma-4-26B-A4B-it-AWQ-4bitservices:
vllm:
image: rocm/vllm:latest
command: vllm serve google/gemma-4-26B-A4B-it --host 0.0.0.0 --port 8000
ports:
- "8000:8000"
ipc: host
devices:
- /dev/kfd
- /dev/dri
group_add:
- video
security_opt:
- seccomp=unconfined
volumes:
- ~/.cache/huggingface:/root/.cache/huggingfaceAMD’s rocm/vllm image. API on port 8000.
sglang serve google/gemma-4-26B-A4B-it --host 0.0.0.0 --port 30000Needs SGLang’s ROCm build (Instinct MI300-class GPUs).
from transformers import pipeline
pipe = pipeline("image-text-to-text", model="google/gemma-4-26B-A4B-it", dtype="auto", device_map="auto")
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
{"type": "text", "text": "Describe this image."},
],
}
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])Vision models that fit the Radeon RX 7900 XTX comfortably
All that fit →| Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|
| 27.8B | Q5_K_M | ≈22 GBFits in memory | ~27 tok/s | |
| 27.8B | Q5_K_M | ≈22 GBFits in memory | ~27 tok/s | |
| 27.4B | Q4_K_M | ≈21 GBFits in memory | ~32 tok/s | |
| 27.8B | Q5_K_M | ≈22 GBFits in memory | ~27 tok/s | |
| 29.8B | Q5_K_M | ≈22 GBFits in memory | ~25 tok/s | |
| 36BA3B | Q4_K_M | ≈22 GBFits in memory | ~290 tok/s |
Other GPUs that run gemma-4-26B-A4B-it
Frequently asked questions
- Can the Radeon RX 7900 XTX run gemma-4-26B-A4B-it?
- Yes. The Radeon RX 7900 XTX runs gemma-4-26B-A4B-it fully in VRAM at Q5_K_M (about 20 GB of 24 GB), at roughly 186 tokens/s.
- How fast does gemma-4-26B-A4B-it run on the Radeon RX 7900 XTX?
- An estimate from memory bandwidth gives about 186 tokens/s at Q5_K_M.
- Which quantisation of gemma-4-26B-A4B-it should I use on the Radeon RX 7900 XTX?
- Q5_K_M: the largest one that fits. Expected quality: near-lossless compared with the full-precision weights.
Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.