Google's Gemma 4 stealth update: 70% faster, fewer bugs, same name

Google dropped a quiet but significant update to its open-source Gemma 4 model family, and the developer community has a bone to pick about how it was handled. The refresh fixes genuinely annoying problems — truncated responses and broken tool calling — while delivering a serious speed boost on Nvidia Hopper GPUs. But the company shipped all of this under the identical Gemma 4 name, skipping a version bump like Gemma 4.1 entirely, which has led to predictable pushback from engineers who need deterministic reproducibility.

Enabling Flash Attention 4 is where the real gains live. Google’s numbers show prompt processing speeding up by 25 to 70 percent, and time to first token dropping by up to 31 percent. That’s not a marginal tweak — it’s the difference between a model that feels sluggish in production and one that snaps. The tool calling bugs that got squashed were equally important. Anyone who’s built agentic workflows knows that a model randomly failing to trigger an API call isn’t a minor annoyance; it breaks chains of reasoning entirely.

For image processing workloads, there’s a practical tip buried in the update. Users can manually bump the max_soft_tokens parameter from 280 to 1,120, which Google says delivers sharper OCR results and supports resolutions up to 2.51 megapixels. The company even stood up an interactive configurator on Hugging Face to make the adjustment easier. While the official benchmarks only compare the 31B and E4B variants against their predecessors, the Hugging Face repository confirms that every parameter size in this generation got refreshed — including the relatively new 12B model.

The naming decision is what’s irking the community, and I get it. When you silently replace a model’s weights without changing its identifier, you create a reproducibility nightmare for anyone who pinned their pipeline to a specific checkpoint. It’s the kind of move that erodes trust in open-source releases, even when the underlying improvements are solid. Google fixed real problems here — the question is whether they’ll learn that versioning discipline matters as much as the engineering.