Instructions to use google/gemma-4-26B-A4B-it with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use google/gemma-4-26B-A4B-it with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="google/gemma-4-26B-A4B-it")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("google/gemma-4-26B-A4B-it")
model = AutoModelForMultimodalLM.from_pretrained("google/gemma-4-26B-A4B-it", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
AMD Developer Cloud
Local Apps Settings

vLLM

How to use google/gemma-4-26B-A4B-it with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "google/gemma-4-26B-A4B-it"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/gemma-4-26B-A4B-it",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/google/gemma-4-26B-A4B-it

SGLang

How to use google/gemma-4-26B-A4B-it with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "google/gemma-4-26B-A4B-it" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/gemma-4-26B-A4B-it",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "google/gemma-4-26B-A4B-it" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/gemma-4-26B-A4B-it",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use google/gemma-4-26B-A4B-it with Docker Model Runner:
```
docker model run hf.co/google/gemma-4-26B-A4B-it
```

GGUF available — Cerebellum v3 (11 GB, ablation-guided mixed-precision)

#37

by deucebucket - opened May 1

Discussion

deucebucket

May 1

Ablation-guided mixed-precision GGUF quant for running this model in llama.cpp / ollama:

deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v3-GGUF — 11 GB

We ran 90 individual ablation experiments to measure which tensors are sensitive vs. tolerant across this MoE architecture and assigned precision accordingly — 99 per-tensor overrides on top of Q3_K_M with bartowski's imatrix.

Benchmarks (RTX 3090):

Benchmark	Cerebellum v3 (11 GB)	Q3_K_M (13 GB)	Q4_K_M (16 GB)
WikiText PPL	19,826	42,369	27,362
HumanEval pass@1	67.1%	62.2%	59.8%
ARC-Challenge	95.5%	95.2%	96.7%
HellaSwag	83.8%	86.6%	85.2%
MMLU-Redux	71.3%	73.7%	72.7%

53% better PPL than Q3_K_M, 28% better than Q4_K_M, at 5 GB smaller. Fits a 24 GB GPU with room for 4K context. Method details in the model card.

thnamratha

Google org May 8

•

edited May 8

Thanks for bringing this to our attention!
Apologies for the late reply.
The technical details on the ablation-guided mixed-precision for Cerebellum v3 are very interesting.
It seems like the link provided for the GGUF model is currently showing a 404 error. Could you please provide an updated or corrected link so we can take a closer look at the model and method details?

deucebucket

May 9

Thanks for bringing this to our attention!
Apologies for the late reply.
The technical details on the ablation-guided mixed-precision for Cerebellum v3 are very interesting.
It seems like the link provided for the GGUF model is currently showing a 404 error. Could you please provide an updated or corrected link so we can take a closer look at the model and method details?

https://huggingface.co/deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF

a couple of changes since then, but hopefully improvements.

deucebucket

Jun 12

•

edited Jun 13

hey, following up since you said you wanted a closer look (and my last comment got truncated mid-table, fixing that here).

current 26B builds, both 11.7 GB, measured on an RTX 3090 with llama.cpp. per-question results are in each repo's benchmark_results folder:

Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF — current file is v6.1 (same tensor allocation as v6, updated chat-template metadata)
Gemma-4-26B-A4B-it-Heretic-Cerebellum-GGUF — same recipe on coder3101's heretic checkpoint, kept separate on purpose

measured scores:

v6/v6.1: ARC-Challenge 95.56, HellaSwag 84.55, MMLU-Redux 71.33
heretic: ARC-Challenge 95.48, HellaSwag 83.49, MMLU-Redux 71.42, HumanEval+ 89.63 (chat harness, thinking off)

the rest of the family is covered too: E4B v2, an E4B heretic, and E2B v2.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment