Gemma: Google's open models with image understanding, from Gemma 3 to Gemma 4

Gemma is Google's open lineup. The previous generation, Gemma 3 (12B and 27B), has been replaced in the Yenisei catalog by Gemma 4 in two sizes, 26B (gemma-4-26b) and 31B (gemma-4-31b). Both understand images and video, can reason and call tools, and have a context of about 260K tokens.

Below is why you would pick an open model when closed flagships exist, and which size to choose.

Why open models at all

The main thing about an open model is predictability and control:

• Stable behavior. An open model doesn't change under your feet: the version is fixed, the behavior is reproducible.
• Price at volume. On high-volume tasks an open model often comes out cheaper than a closed flagship.
• Clear limitations. You know exactly what you are working with, with no black box.

26B versus 31B

Gemma 4 26B is lighter and cheaper. Good for streaming tasks: classification, extraction, simple answers, "what is in this picture" image analysis.

Gemma 4 31B is stronger on tasks that need a more coherent answer.

The rule: start with 26B. If the quality isn't there, move to 31B. Then, if that isn't enough either, move to a larger model such as Kimi K3 (kimi-k3) or Yenisei Max (yenisei-max).

Example with an image

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-31b",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "How many people are in the photo and what are they doing?"},
        {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}
      ]
    }]
  }'

Where Gemma doesn't belong

On complex reasoning it falls short of the largest models. Gemma is chosen for its low price, stability, and good quality on high-volume tasks. If the task is hard, move to a larger model. Payment is in rubles, 1 credit = 1 ₽.