T5Gemma 2
2025-12-18 · Publisher : Google · Domain : Artificial Intelligence
Google releases T5Gemma 2, the new generation of its encoder-decoder models based on Gemma 3: the first multimodal and long-context encoder-decoder models, in three compact sizes.
What changes
- Multimodality: image understanding in addition to text, via an efficient vision encoder.
- Extended context up to 128,000 tokens thanks to Gemma 3's alternating local and global attention.
- More than 140 languages supported out of the box.
- Architecture changes: tied embeddings between encoder and decoder, merged decoder attention.
### Specifications - Pre-trained sizes: 270M-270M (about 370M in total, excluding the vision encoder), 1B-1B (about 1.7B), 4B-4B (about 7B). - Checkpoints available on Kaggle, Hugging Face, Google Colab and Vertex AI. ### Pricing Open models, with no usage pricing of their own. ### Benchmarks The blog post presents results against reference models as charts, with no figures in the text. ### Key takeaways Extends the encoder-decoder family to multimodal and long-context uses, with small models suited to devices.
Links
Tags : Open Source, LLM, Google, Gemma, Encodeur-décodeur