Google Introduces T5Gemma 2: Multimodal, Long-Context AI Models

Details

• Google has launched T5Gemma 2, a new line of encoder-decoder large language models that incorporates multimodal and long-context processing, leveraging advances from the Gemma 3 architecture.
• The family features three compact model sizes—270M-270M, 1B-1B, and 4B-4B—and is offered as open-weight checkpoints on Hugging Face, Kaggle, Colab, and Google Vertex AI.
• T5Gemma 2 introduces innovations such as tied embeddings between encoder and decoder, as well as merged self- and cross-attention in the decoder, resulting in fewer parameters and greater inference efficiency.
• The models support over 140 languages, feature a 128,000-token context window, and can process both images and text, enabling sophisticated visual question answering and reasoning tasks.
• Benchmarks show T5Gemma 2 surpasses both its predecessor and Gemma 3 in multimodal, long-context, coding, reasoning, and multilingual capabilities, making it a strong open alternative for diverse applications.

Impact

With T5Gemma 2, Google advances the democratization of powerful vision-language AI, providing high performance in compact, accessible models. The launch helps close the gap with leading proprietary systems and targets developers seeking efficient solutions for specialized, on-device, or research-focused deployments. Google’s strategy emphasizes pushing encoder-decoder innovation in a space dominated by larger, less efficient models.