.svg){kind=spacer}
Details
- Google AI Developers have unveiled T5Gemma 2, the successor to Gemma 3, on December 18, 2025.
- This new model features native multimodality, enabling it to handle combined text and image inputs and outputs without the need for separate vision encoders.
- T5Gemma 2 expands its context window to process far larger documents than Gemma 1's previous 32,000-token limit, though the new maximum was not specified.
- The tokenizer now accommodates over 140 languages, aiming for improved translation accuracy and code-switching for global users.
- With a refreshed architecture, the model employs sparsely gated Mixture-of-Experts layers and improved attention routing, resulting in better throughput on TPUs and the latest Nvidia GPUs.
- Pre-trained checkpoints are immediately available on Kaggle and Hugging Face, with Colab notebooks for demonstrations and Vertex AI enabling managed inference with a single click.
- All assets are released under a permissive open-weights license, supporting commercial fine-tuning, evaluation, and experimentation on devices.
Impact
Reviving the encoder-decoder line positions Google competitively against Meta’s Llama 3 and OpenAI’s GPT-4o, contributing to a broader ecosystem of openly accessible models. Integrated multimodal features can lower costs for startups by simplifying product stacks, while the open licensing makes T5Gemma 2 attractive for enterprises navigating new AI compliance standards. Its extensive language support may spur global adoption and increased development of multilingual applications, especially in emerging markets facing language barriers.
