Gemma 3
2025-03-12 · Publisher : Google · Domain : Artificial Intelligence
Google DeepMind releases Gemma 3, its new family of open models based on Gemini 2.0 technology: four sizes from 1 to 27 billion parameters, multimodal, with a 128,000-token context.
What changes
- Four sizes: 1B, 4B, 12B and 27B, in pre-trained and instruction-tuned versions.
- Vision: the 4B, 12B and 27B versions accept images and short videos (SigLIP encoder), 1B is text-only.
- Support for more than 140 languages in pre-training, including more than 35 ready to use out of the box.
- Function calling and structured outputs; 27B ranked at 1,338 Elo on Chatbot Arena with a single GPU, according to Google.
- Launch of ShieldGemma 2, a 4-billion-parameter image safety classifier (dangerous, sexually explicit, violent content).
### Specifications - Context: 128,000 tokens. - Modalities: text and image (except 1B), text output. - Availability: Google AI Studio, Hugging Face, Kaggle, Ollama, Vertex AI, Cloud Run. - API identifier for Gemma 3 27B: `gemma-3-27b-it`. ### Pricing Open-weights model, with no usage price of its own. Access to gemma-3-27b-it in the Gemini API is subject to AI Studio quotas; price not disclosed at announcement. ### Benchmarks | Metric | Score | |---|---| | Chatbot Arena, Gemma 3 27B (Elo) | 1,338 | Score published by the vendor. The post states that Gemma 3 27B outperforms comparable models that require up to 32 GPUs. ### Key takeaways Gemma 3 becomes a solid foundation for local or sovereign deployments of multimodal models, with a low hardware floor (one GPU for the 27B size).
Links
Tags : Open Source, LLM, Google, Gemma, Multimodal