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。
“B端客户不在乎定不定制，他在乎的是ROI测算和安全性。”朱健楠说。这意味着产品打磨的逻辑必须改变，“不是为一个客户开一套模，但产品功能每年起码要迭代一次，避免客户不买单的风险。我们会把80%的共性需求收回来，适配80%的场景，剩下的20%靠功能模块补齐。“
，详情可参考新收录的资料
know that that's UTF-8. That's much more complicated to do in hex.
"We've had to very quickly become keenly aware of the fact that we are being watched in a really intense and new way by the government and their partners and Big Tech," said Hunt. "And that collusion is out in the open."，更多细节参见新收录的资料
Add Entrepreneur
1. List all open issues: gh issue list --state open --json number。业内人士推荐新收录的资料作为进阶阅读