selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

google logo

Can the Instinct MI50 32GB run gemma-4-26B-A4B-it?

gemma-4-26B-A4B-it·25.8B (4B active)·Instinct MI50 32GB·32 GB VRAM

Saved in this browser, used on every model page.
Yes

Yes. The Instinct MI50 32GB runs gemma-4-26B-A4B-it fully in VRAM at Q8_0 (about 29 GB of 32 GB), at roughly 133 tokens/s.

best quant
Q8_0
needs
29 GB
usable
32 GB
est. speed
~133 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 Instinct MI50 32GB

QuantQualityMemoryFitSpeed
Q2_K Heavy loss11 GBfits~428 tok/s
Q3_K_M Noticeable loss15 GBfits~288 tok/s
Q4_K_M Small loss18 GBfits~232 tok/s
Q5_K_M Near-lossless20 GBfits~198 tok/s
Q6_K Near-lossless23 GBfits~171 tok/s
Q8_0 pickLossless29 GBfits~133 tok/s
FP16 / BF16 Lossless52 GBoffload~9.9 tok/s

Speed is a bandwidth-based estimate for generating text (1024 GB/s); prompt processing and long contexts are slower.

Measured speed on the Instinct MI50 32GB

No community reports for this pair yet.

Report your own speed →

Run gemma-4-26B-A4B-it on the Instinct MI50 32GB

Using Q8_0, the best fit for Instinct MI50 32GB.

Ollama · From the Ollama librarydocs
ollama run gemma4:26b

The library build picks its own default quantisation. Ollama’s ROCm/CUDA builds don’t cover this GPU. Recent Ollama releases use Vulkan when its Vulkan backend and driver are installed (older ones needed OLLAMA_VULKAN=1); llama.cpp or LM Studio with Vulkan are the safer bet.

Ollama · Q8_0 straight from Hugging Facedocs
ollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0

Ollama’s ROCm/CUDA builds don’t cover this GPU. Recent Ollama releases use Vulkan when its Vulkan backend and driver are installed (older ones needed OLLAMA_VULKAN=1); llama.cpp or LM Studio with Vulkan are the safer bet.

llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0

Needs a Vulkan build of llama.cpp. Downloads the GGUF on first run.

llama.cpp · Docker with Vulkandocs
docker run -d --name llama --device /dev/dri -p 8080:8080 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-vulkan \
  -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 --host 0.0.0.0 --port 8080 -ngl 99

Server and web UI on port 8080.

LM Studio · Download with the lms CLIdocs
lms get unsloth/gemma-4-26B-A4B-it-GGUF@Q8_0

Then load it from the LM Studio app or with `lms load`.

Transformers · Ask about an imagedocs
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 Instinct MI50 32GB comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
27.8BQ6_K≈25 GBFits in memory~25 tok/sCommercial OK
27.8BQ6_K≈25 GBFits in memory~25 tok/sCommercial OK
31.3BQ5_K_M≈29 GBFits in memory~25 tok/sCommercial OK
27.4BQ6_K≈26 GBFits in memory~25 tok/sConditions
27.8BQ6_K≈24 GBFits in memory~25 tok/sCommercial OK
meta-models logoMuse-Glimmer-30Bmeta-models
29.8BQ8_0≈30 GBFits in memory~18 tok/sCommercial OK

Other GPUs that run gemma-4-26B-A4B-it

Frequently asked questions

Can the Instinct MI50 32GB run gemma-4-26B-A4B-it?
Yes. The Instinct MI50 32GB runs gemma-4-26B-A4B-it fully in VRAM at Q8_0 (about 29 GB of 32 GB), at roughly 133 tokens/s.
How fast does gemma-4-26B-A4B-it run on the Instinct MI50 32GB?
An estimate from memory bandwidth gives about 133 tokens/s at Q8_0.
Which quantisation of gemma-4-26B-A4B-it should I use on the Instinct MI50 32GB?
Q8_0: the largest one that fits. Expected quality: 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.