Google DeepMind Releases Gemma 4 Technical Report on Multimodal Models
The new report describes open-weight multimodal models from 2.3B to 31B parameters, and it quickly sent X into a side debate about how much training data Gemma 4 may have used.
401.1K12140879.2K
“Happy to share we just published Gemma 4 technical report! Take a look”
Olivier Sanseviero@osansevieroTECH#889Story Brief
Google DeepMind researchers Olivier Bachem, Petar Veličković and Olivier Sanseviero posted that the Gemma 4 Technical Report is now live, describing a family of open-weight, natively multimodal models from 2.3B to 31B parameters. According to those posts and the paper title and abstract on the linked arXiv page, the report covers dense and mixture-of-experts models, a 12B encoder-free variant for raw audio, and benchmark claims on STEM, multimodal and long-context tasks.
“seems like gemma 4 was trained on MUCH more tokens than the previous iteration”
Elie Bakouch@eliebakouch
