Gemma 4 静默更新：修复工具调用错误并提升 Hopper GPU 性能

谷歌发布了其开源 AI 模型 Gemma 4 的更新，该更新加快了在 Nvidia Hopper GPU 上的性能，修复了工具调用错误，并解决了回答被截断的问题。谷歌表示，开启 Flash Attention 4 可以将模型处理输入提示的速度提升 25% 到 70%。首次生成令牌的时间最多可减少 31%。谷歌还修复了工具调用功能中的错误，该功能允许模型自行触发外部工具。

谷歌表示，还减少了模型回答被截断或返回不完整响应的情况。对于图像处理，用户可以手动将 "max_soft_tokens" 参数从 280 提高到 1,120，以获得更清晰的 OCR 结果，并支持最高 2.51 兆像素的分辨率。谷歌在 Hugging Face 上提供了一个互动配置器：https://huggingface.co/spaces/google/gemma4_vision_token_budget。已发布的基准测试仅比较了 31B 和 E4B 版本与其前代模型的性能，但 Hugging Face 仓库：https://huggingface.co/collections/google/gemma-4 显示该代模型所有参数大小都已更新，包括最新的 12B 版本：https://the-decoder.com/google-deepminds-gemma-4-12b-squeezes-multimodal-ai-on-to-a-laptop-with-just-16-gb-of-ram/。社区对谷歌以相同的 "Gemma 4" 名称发布更新表示不满，而不是将其标记为单独的版本，如 "Gemma 4.1"：https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/。

保持 AI 动态。内容清晰、有用，无废话。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Google shipped an update to its open AI model Gemma 4 that speeds up performance on Nvidia Hopper GPUs, fixes tool calling bugs, and addresses problems with truncated responses. Turning on Flash Attention 4 boosts the speed at which the model processes incoming prompts by 25 to 70 percent, according to Google. Time to first token drops by up to 31 percent. Google also fixed bugs in tool calling, the feature that lets the model trigger external tools on its own.

Google says it also cut down on cases where the model would cut answers short or return incomplete responses. For image processing, users can manually raise the "max_soft_tokens" parameter from 280 to 1,120 to get sharper OCR results and support resolutions up to 2.51 megapixels. Google put up an interactive configurator on Hugging Face：https://huggingface.co/spaces/google/gemma4_vision_token_budget for that. The published benchmarks only compare the 31B and E4B variants against their predecessors, but the Hugging Face repository：https://huggingface.co/collections/google/gemma-4 shows that all parameter sizes in this model generation got updated, including the newest 12B release：https://the-decoder.com/google-deepminds-gemma-4-12b-squeezes-multimodal-ai-onto-a-laptop-with-just-16-gb-of-ram/. The community has pushed back on Google shipping the update under the same "Gemma 4" name：https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/ instead of tagging it as a separate version like "Gemma 4.1."

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/