I was fortunate to be invited to the Gemma Developer Day at Google Paris, where the Gemma 3 family of open-weight models was officially launched. 🚀
The 1b is a text-only model (no vision encoder) focused on English and a great candidate for integration on your favourite edge device – running on a single GPU. The 4b, 12b and 27b models support over 140 languages and are multi-modal. Further, Google provides quantized versions of the Gemma 3 models in three weight representations: per-channel int4, per-block int4, and switched fp8. They’ve also significantly extended the context window from Gemma 2 (8k) to 128k.
While the benchmark results certainly look impressive, with the instruction-tuned 27b model achieving a great ranking on the Chatbot Arena as well (currently 1339), the proof will be in pudding! Looking forward to taking these open models for a spin and see what they can do on a realistic use-case.