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
```

Is base a direct parent of `it`?

#57

by olisicky - opened 4 days ago

Discussion

olisicky

4 days ago

Thank you for gemma4 models, they are amazing! I have a question, though.

We measured weight-space distance between base and instruct releases across four model pairs. Llama-3.1-8B and Qwen3.5-35B-A3B both come out at cosine ~0.999 with weight norms shrinking ~0.4%. Both Gemma 4 pairs — the dense 31B and the 26B-A4B MoE — come out at cosine 0.81–0.84 with weight norms growing 1.2–2.1×, and with q_norm/k_norm bit-identical between the two releases.

Is the released base checkpoint the direct parent of the it checkpoint, or is there an unreleased stage in between? And relatedly: for adding a new language to Gemma 4, is SFT on the it checkpoint the intended path, given that (it - base) doesn't look like a usable task vector?

Thank you very much,

Ondřej

pannaga10

Google org 3 days ago

Hi @olisicky
Gemma4 instruction-tuned models as being obtained from the corresponding pre-trained (PT/base) models through a post-training pipeline, similar to Gemma 3. The intended flow is PT/base → post-training → IT. The IT checkpoint includes the instruction-tuning/post-training stage, including the additional training for thinking mode, safety, factuality, attribution, and refusals.
For adding a new language, we recommend starting from the instruction-tuned checkpoint if you want to retain the model's existing instruction-following behavior while adapting it to the new language. The PT/base checkpoint is more appropriate for continued pre-training or further model development.
You can start from the -it checkpoint and fine-tune it for the target language/task.
For ref :: https://ai.google.dev/gemma/docs/tune

Thanks

olisicky

3 days ago

hi @pannaga10

thanks for reply. I tested also gemma3 model which shows similar small drift (as llama31, qwen3.5) from pt version and works well for task vector extraction etc. The big difference for gemma-4-26B-A4B and dense version is interesting, though. Have you tried some merge techniques or something similar, please?

Thanks

Ondřej

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