Gemma3

Gemma 3 is Google DeepMind's state-of-the-art, open-weight model family: it delivers multimodal (text/image) and 140-language support, running efficiently on a single GPU.

This is the next generation of portable AI: Gemma 3, built on Gemini 2.0 technology, is a collection of open-weight models (270M to 27B parameters) engineered for on-device deployment. The 27B variant, for instance, outperforms larger models like Llama3-405B in human preference evaluations, yet requires only a single GPU or TPU. Key features include a 128K-token context window and unparalleled multilingual capability, supporting over 140 languages for global application development. We’ve made high-performance, multimodal AI accessible: get it running on your laptop, workstation, or smartphone.

https://deepmind.google/models/gemma/