Gemma 4 Vision

A lot of people in the [Gemma 4 Model Request Thread](https://www.reddit.com/r/LocalLLaMA/comments/1srgqk4/which_gemma_model_do_you_want_next/) were asking for better vision capabilities in the next Gemma Model. This tells me that people are not configuring Gemma 4's vision budget. Gemma 4 ships with [Variable Image Resolution](https://huggingface.co/google/gemma-4-31B-it#5-variable-image-resolution). The default max vision budget is 280 ([~645K pixels](https://huggingface.co/docs/transformers/model_doc/gemma4)) which is way too less. In this mode, it fails to OCR tiny details. It's essentially blind in my books. In llama.cpp, you can configure Gemma 4's vision budget with 2 parameters --image-min-tokens and --image-max-tokens. The engine will try to fit the image within those bounds. I believe the default is 40 and 280 respectively. This is Gemma 4's default from Google's side but it's way too low. I like to run them at 560 and 2240 respectively and it's able to pick up very minute and hazy details within images. Why 2240 - isn't that double of the max from Google (1120)? In my testing, 2240 for some reason works better than 1120. I suspect this might be because of llama.cpp's implementation where it tries to fit the image between min and max tokens. Additionally, you will also have to set --batch-size and--ubatch-size above whatever value you choose for image-max-tokens. I run them at 4096 (for --image-max-tokens 2240). This will consume a lot more VRAM (63 GB (default) to 77 GB (4096 batch) for q8_0 at max context). If you use Ollama, you are likely SOL until and if they care to fix [this](https://github.com/ollama/ollama/issues/15626). It's worth it though, with a higher vision budget, Gemma 4 is pretty much SOTA for Vision and pretty much destroys anything else especially for OCR - Qwen 3.5, Qwen 3.6, GLM OCR (or any other random OCR), Kimi K2.5. I haven't tested Kimi K2.6 and I refuse to touch Cloud Models.