Google Ships Gemma 4 Stealth Update Fixing Tool Calling and Truncation Bugs
Tags AI · OSS · Developer Tools
Google released an unversioned update to its open Gemma 4 model family on July 16, 2026, adding Flash Attention 4 support (25-70% faster prompt processing, up to 31% lower time-to-first-token), fixing tool-calling bugs, and reducing truncated responses. The 31B variant shows up to 10.1% improvement in agentic reasoning benchmarks. Image processing now supports up to 2.51 megapixels via a raised max_soft_tokens parameter. Community criticism noted the lack of version bump (e.g., Gemma 4.1).
Technical significance
Flash Attention 4 integration delivers meaningful inference speedups on Hopper GPUs without model architecture changes — a pure software optimization. The tool-calling fixes address a critical gap for agentic workflows. However, shipping under the same Gemma 4 name breaks reproducibility expectations for developers pinning model versions, and the absence of benchmark data for all parameter sizes (only 31B and E4B published) limits upgrade confidence.