Google DeepMind has announced T5Gemma, a new collection of encoder-decoder large language models. The name points to a blend of the T5-style encoder-decoder design with the Gemma model family, which previously centered on decoder-only architectures.
According to the announcement, T5Gemma is positioned as an addition to the Gemma lineup rather than a replacement. Because the source is a single introductory post, there are no independent comparisons or performance details to weigh; the significance lies in the architectural expansion of the Gemma family.
The release signals continued investment in encoder-decoder models, which process input and output separately, as an alternative to the dominant decoder-only approach. Further technical specifications and use cases are expected to follow.