Google updates Gemma 4 without a new version name

Google has updated its open Gemma 4 AI model line under the same Gemma 4 name, saying the refresh improves Nvidia Hopper GPU performance, tool calling and incomplete-answer behavior. The decision to ship the changes without a distinct version label has drawn pushback from parts of the developer community because it makes model provenance harder to track.

Google said enabling Flash Attention 4 increases prompt processing speed by 25% to 70% on Nvidia Hopper GPUs. The company also said time to first token falls by as much as 31%, a metric that matters for latency-sensitive products where model response time can be the difference between a usable agent and a slow demo.

The update also targets tool calling, the mechanism that lets a model invoke external software, APIs or functions as part of a task. Google said it fixed bugs in that area and reduced cases where Gemma 4 would stop early or produce incomplete responses. It did not disclose how frequently those failures occurred before the update.

Benchmarks show uneven but positive gains

Google’s published benchmark data focuses on the 31B and E4B variants compared with earlier versions. For Gemma 4 31B, the company reported gains across tested tool-use and agentic reasoning scenarios: BFCL reached 74.2%, up 0.4 points; TB2 reached 25.8%, up 4.5 points; Tau2 Retail reached 77.6%, up 3.1 points; Airline reached 84%, up 2.0 points; and Telecom reached 62.7%, up 10.1 points.

The telecom result is the largest improvement cited in the disclosed benchmark image. Google did not provide broader third-party validation in the material described, and the benchmark comparison does not cover every model size in the same detail.

For vision and document workflows, Google said users can manually raise the max_soft_tokens parameter from 280 to 1,120. The company said that can improve OCR sharpness and allow image resolutions up to 2.51 megapixels. Google also published an interactive configurator on Hugging Face for adjusting that vision token budget.

Same name, changed weights

The Hugging Face repository for Gemma 4 shows that the refresh applies across all parameter sizes in this model generation, including the newer 12B release. Google’s public benchmark discussion, however, only compares the 31B and E4B versions with their predecessors.

That gap between broad repository updates and narrower published comparisons is part of why naming matters. Developers and companies evaluating open-weight models often pin exact versions for testing, compliance reviews and regression analysis. If changed weights ship under the same product name, teams have to rely more heavily on repository metadata and hashes to know what they are running.

Some community members have criticized Google for not labeling the update as a separate release such as Gemma 4.1. Google has not disclosed a new version number for the refresh in the details cited here.

This story draws on original reporting from The Decoder.