Google has introduced T5Gemma 2, the next evolution of its encoder-decoder model family built on the powerful Gemma 3 architecture, bringing significant architectural innovations and expanded capabilities to AI research and development. Unlike traditional decoder-only models, T5Gemma 2 uses a classic encoder-decoder design with tied word embeddings and merged attention mechanisms to reduce parameters and improve efficiency, making it suitable for rapid experimentation and deployment across a range of environments. The new models are available in compact configurations—from millions to billions of parameters—and natively support multimodal inputs, enabling them to process both text and images together for tasks like visual question answering and reasoning.
Also Read: IonQ and KISTI Advance South Korea’s Quantum Ambitions With 100-Qubit System Deployment
T5Gemma 2 also offers ultra-long context handling, with support for sequences up to 128,000 tokens, and robust multilingual understanding across more than 140 languages, inheriting the advanced features of Gemma 3 while outperforming its predecessors on long‐context and multimodal benchmarks. Pre-trained checkpoints are now available to developers, who can further post-train the models for specific applications, underscoring Google’s commitment to open, efficient AI innovation.