T5Gemma 2 is the next evolution of our encoder-decoder family based on Gemma 3, featuring the first multi-modal and long-context encoder-decoder models.
Unlike T5Gemma, T5Gemma 2 adopts tied word embeddings (over encoder and decoder) and merged decoder self- and cross-attention to save model parameters. It offers compact pre-trained models at sizes of 270M-270M (~370M total, excluding vision encoder), 1B-1B (~1.7B) and 4B-4B (~7B) parameters, making them ideal for rapid experimentation and deployment in on-device applications.
Background
With the original T5Gemma, we demonstrated that we could successfully adapt modern, pre-trained decoder-only models into an encoder-decoder architecture, unlocking new versatility. By initializing with weights from a powerful decoder-only model and then applying continued pre-training, we created high-quality, inference-efficient models while bypassing the computational cost of training from scratch.
T5Gemma 2 extends this into the realm of vision-language models by incorporating key innovations from Gemma 3.
What’s new
T5Gemma 2 is more than a re-training. It incorporates significant architectural changes while inheriting many of the powerful, next-generation features of the Gemma 3 family.
Architectural innovations for efficiency
To maximize efficiency at smaller scales, we have introduced key structural refinements:
Tied embeddings: We now tie the embeddings between the encoder and decoder. This significantly reduces the overall parameter count, allowing us to pack more active capabilities into the same memory footprint — crucial for our new compact 270M-270M model.
... continue reading