我在《把离线AI智能体装进口袋里》（The Dawn of Offline AI Agents in Your Pocket）一文中对此进行了详细阐述。但文章中的示例更像是Demo，而非生产解决方案。像 Gemma 3n 这样的模型虽然能够很好地处理函数调用，但它们体积过大：无法集成到应用程序包中，需要单独下载，即使在旗舰机型上推理速度也很慢。在低端设备上，它们根本无法运行。而较小的型号则经常出现故障，难以记住工具。
如果你也正站在孩子入园的门槛前，我想分享几点心得：
。旺商聊官方下载对此有专业解读
Crucially, this distribution of border points is agnostic of routing speed profiles. It’s based only on whether a road is passable or not. This means the same set of clusters and border points can be used for all car routing profiles (default, shortest, fuel-efficient) and all bicycle profiles (default, prefer flat terrain, etc.). Only the travel time/cost values of the shortcuts between these points change based on the profile. This is a massive factor in keeping storage down – map data only increased by about 0.5% per profile to store this HH-Routing structure!
短短四年间，公司完成从设计到制造、从材料到封测的全链条卡位，布局半导体的决心可见一斑。