Gemma 4 Vision

A lot of people in the Gemma 4 Model Request Thread 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. The default max vision budget is 280 (~645K pixels) 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