Gemma 3 brings multimodal AI closer to the product edge

Gemma 3 makes a practical bet: more capable multimodal work can live closer to the product boundary. The family spans 1B to 27B parameters, adds vision understanding, and gives the 4B, 12B, and 27B sizes at least 128K tokens of context. Its report also changes the attention mix to reduce KV-cache growth as context grows. 1

A product can take text, screenshots, scanned pages, and longer source material through one model surface. Google positions Gemma 3 for a single GPU or TPU, with 4B, 12B, and 27B variants supporting vision. Treat 128K as capacity, not a promise of perfect recall. 2 3

Start with google/gemma-3-4b-it on Hugging Face. The model card documents text and image input, text output, transformers >= 4.50.0, and an image-text-to-text pipeline. A sensible first prototype is screenshot or document understanding. Test on your own files, add grounding and citations around the answer, then decide whether retrieval is still needed. 4