Google quietly updated Gemma 4 to fix tool-calling bugs and add Flash Attention 4 on Hopper GPUs
Google shipped a silent update to Gemma 4 fixing tool-calling bugs, truncated responses, and enabling Flash Attention 4 on Hopper GPUs for a reported 25–70% prefill speedup.
Google updated its Gemma 4 model family to fix tool-calling bugs, reduce truncated responses, and enable Flash Attention 4 on Nvidia Hopper GPUs. The update shipped under the same “Gemma 4” name with no version bump, which the community has already pushed back on.
The Flash Attention 4 path is the headline number for anyone running this on Hopper hardware: Google reports a 25 to 70 percent improvement in prompt processing speed and up to 31 percent lower time-to-first-token. The tool-calling fix is the more immediately practical change for agentic use. Google provided updated Jinja chat templates for the 31B, 27B, E4B, and E2B variants on Hugging Face; if you’re running older GGUFs without pulling the new templates, you’re still on the buggy behavior. The fix can be applied in llama.cpp with --chat-template-file pointing at the updated .jinja file. A reasoning budget fix was also merged into upstream llama.cpp (PR #21697). Separately, raising max_soft_tokens from 280 to 1,120 reportedly improves OCR results and unlocks support for higher-resolution images.
The no-version-bump approach is a real problem for reproducibility. If you’re comparing Gemma 4 results from before and after this update, you’re not comparing the same model, and nothing in the name tells you that. It’s the kind of thing that would get called out in any serious benchmark context. Pull the updated templates now if you’re running any Gemma 4 variant for tool use.