Gemma 3 (1B, 4B, 12B, 27B)
Gemma 3: 1B to 27B open models with vision input, 128k context and 140-language support; Google says the 27B rivals much larger models on LMArena.
Adds image and short-video understanding, a 128k-token window, function calling and structured output, and official quantized checkpoints. Designed to fit one GPU or TPU host. ShieldGemma 2 image safety classifier released alongside.
- Date
- Wednesday, 12 March 2025
- Lab
- Google DeepMind
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Context window | 128k tokens 4B, 12B, 27B sizes | company |
| Gemma downloads at first birthday (Feb 2025) | 100M+ 60,000+ community variants | company |
LMArena claim is a preliminary human-preference comparison reported by Google. Weights are under the Gemma terms of use, not an OSI license. Gemma 4 later moved to a new generation (2026-04-02).
Sources
This record was checked against its sources on 6 October 2026. How we check