Google updated Gemma 4 silently — is this the future of open-source maintenance?
google silently updated gemma 4, fixing tool-calling bugs and boosting gpu speed up to 70% via flash attention 4. this forces developers to re-evaluate the frozen model paradigm.
google released a stealth update for the gemma 4 family, activating flash attention 4 on nvidia hopper gpus to reduce time-to-first-token by up to 31% and boost processing speeds by 25-70%. the patch resolves tool-calling failures and truncation issues,提升了电信代理场景的推理能力10.1%, while increasing image resolution support to 2.51 megapixels. unlike standard industry practice, google shipped this under the 'gemme 4' tag rather than 'gemme 4.1', affecting all parameter sizes including the 12b release. this decision has sparked community backlash over model reproducibility and complicates dependency management for downstream developers.