gemini-4IMAGE
Gemini API calls local Whisper instead of translating audio itself
A developer using Gemini 3.8 Flash with an agent found the model called a local Whisper model instead of using its own multimodal translation, and spent nearly 30 minutes thinking.
Zane
@zane_osArchived 18 days agoOriginally posted 2026.10.09
Zane @zane_osOriginally posted 2026.10.09
今天使用 Gemini API 接agent时 发现哈基米已经懒到一定程度了 我想用让哈基米帮我翻译一段音频时 它居然调用本地的Wishper模型 不用它自己的多模态去翻译？！ 草 我用的是 Gemini 3.8 Flash 我还以为正在灰度测试Gemini 4 Argon 路由到新模型上了呢 一个Flash模型思考了快30分钟 害我白高兴一场😡
x.com/zane_os/status/2108536613601267801Archived 2026.10.10 10:01