Gemma 4 vision budgets confuse users

The feature exists, but it’s split across stacks in a way that makes the model feel harder to use than it should.

• –In Transformers, Gemma 4’s image processor accepts `max_soft_tokens`, with supported budgets of 70, 140, 280, 560, and 1120.
• –In vLLM, the default is set at launch with `--mm-processor-kwargs '{"max_soft_tokens": 560}'`, and it can be overridden per request.
• –Higher budgets are the right choice for OCR, document parsing, handwriting, and small text; lower budgets favor captioning, screenshots, and faster multimodal work.
• –The reddit post is really surfacing a documentation problem: the advice is there, but the configuration path is buried under model, processor, and server-specific APIs.
• –Audio is a separate constraint path, not the same budget knob: it only applies on E2B/E4B and is capped at 30 seconds.