Google recently released T5Gemma 2, a next-generation encoder-decoder model based on the Gemma 3 architecture. Compared to its predecessor, T5Gemma 2 introduces multiple architectural innovations, including tied word embeddings and merged attention mechanisms, significantly reducing model parameters. The new model supports multimodal processing capabilities, able to simultaneously understand and process images and text; its context window expands to 128K tokens, greatly enhancing long text processing capabilities; and it supports over 140 languages, possessing powerful multilingual processing abilities. Performance tests show that T5Gemma 2 surpasses its predecessor in multimodal, long-context, encoding, and reasoning tasks. The series offers three scales of pre-trained models: 270M-270M, 1B-1B, and 4B-4B, suitable for on-device applications and downstream task development. The models are now available for download on platforms such as Kaggle and Hugging Face, providing AI researchers and developers with a powerful new tool.
