Google released a new version of its Gemma 4 open source model that improves performance on Nvidia Hopper GPUs while correcting specific technical errors. Activating Flash Attention 4 increases prompt processing speed by between 25 and 70 percent, with time to first token dropping by up to 31 percent. The update also resolves bugs in tool calling, a feature allowing the model to trigger external applications independently, and reduces instances of truncated or incomplete answers. For image processing tasks, users can now manually increase the max_soft_tokens parameter from 280 to 1,120 on the Hugging Face configurator. This change enables sharper OCR results and supports input resolutions up to 2.51 megapixels.
The decision to ship these improvements under the same Gemma 4 name rather than creating a distinct version number has drawn criticism from the developer community. Although official benchmarks only compare the 31B and E4B variants against previous models, internal repository data confirms that all parameter sizes, including the 12B release, received the update. This approach prioritises rapid deployment over clear version tracking, which may complicate reproducibility for researchers relying on specific model snapshots.
- Flash Attention 4 boosts speed by 25 to 70 percent
- Max soft tokens increase to 1,120 for better OCR
- Community objects to lack of version number change
