Back

#Google releases #Gemma3 with support for 140+ languages, image & video input, 128k context window, and top #LMArena score of 1338 at just 27B parameters. Available via #AIStudio, #HuggingFace, #Ollama, #Vertex, and more.

Model outperforms larger competitors with ELO score of 1338 while maintaining compact 27B size. Available in four sizes (1B, 4B, 12B, 27B) as pre-trained and instruction-tuned versions. Multimodal capabilities support vision-language tasks with SigLIP encoder.

Gemma3 by #Google Brings Advanced Capabilities to Open Models

The latest version of #Gemma brings significant improvements to the open model ecosystem

Supports 140+ languages with improved multilingual tokenizer
Multimodal capabilities with integrated SigLIP vision encoder for image and video input
Enhanced reasoning, math, and coding with RLHF, RLMF, and RLEF training

Expanded context window of 128k tokens for complex tasks
Achieves 1338 LMArena score, making it the top compact open model
Available in 1B, 4B, 12B, and 27B parameter sizes
Compatible with Gemma 2 dialog format for easy integration

Pre-trained models trained on 2-14T tokens using Google TPUs and JAX
Includes ShieldGemma 2, a specialized 4B image safety classifier

Download today via AI Studio, Hugging Face, Ollama, Vertex AI, or your favorite OS tools.

https://developers.googleblog.com/en/introducing-gemma3/