Google Space demos Gemma 4's tunable vision token budget

Google has quietly put up a Hugging Face Space that isn't a benchmark leaderboard or a marketing microsite. It's a knob. You feed the Space an image, then choose how many tokens Gemma 4 spends encoding it: 70, 140, 280, 560, or 1120. The whole point is to watch, on your own image, what you gain and what you lose as the budget moves.

That is a more useful demo than it sounds. According to Google's Gemma 4 launch post on Hugging Face, the new encoder "preserves the original aspect ratios and can encode images to a few different token budgets (70, 140, 280, 560, 1120)." That is the departure from the usual VLM habit of squashing every image into a fixed square like 224x224. The Gemma 4 model card is explicit about how to spend the budget: lower budgets suit classification tasks, higher budgets work better for OCR, document parsing, or reading small text.

The context for why the dial matters is the rest of the release. Google shipped Gemma 4 on April 2, 2026 under Apache 2.0 across five sizes, per the launch post: an E2B and E4B pair aimed at on-device use, a 12B Unified encoder-free variant, a 26B A4B mixture-of-experts with 3.8B active parameters, and a 31B dense flagship. Google reports the 31B achieves 76.9% on MMMU Pro and 85.6% on MATH-Vision, and the models natively respond in JSON with detected bounding boxes with no need for grammar-constrained generation. Take those numbers as reported, not settled.

The honest caveat is that the Space itself doesn't tell you how latency and quality actually trade off on your hardware and your images. Google's guidance is directional, and the retrieved posts don't publish per-budget accuracy curves or third-party benchmarks. If your product depends on OCR accuracy at 70 tokens versus 1120, you will need to measure it yourself, which is presumably why this Space exists.

The upside, if the reported numbers hold, is that a team building document parsing, UI understanding, or edge-side visual agents now has an Apache 2.0 model where the compute bill scales with the task instead of with a fixed image tokenizer someone else picked.