MiniCPM 8B 成为老挝国家模型,端侧小模型加速落地多国
面壁智能转发并认同 @xiaoanlab 对 MiniCPM 端侧落地的观察:截至 2026 年 11 月 MiniCPM 下载量超过 7000 万,其 8B 老挝国家模型接入老挝政府 G-Chat v2 通讯软件,老挝翻译在国际基准上超过 GPT-4o。马来西亚用 VoxCPM2 做语音交互,肯尼亚一家银行在其上运行端侧风控 Agent,开发者基于它推出离线应用;小模型之争正从参数量转向谁真正嵌入日常系统。
引用内容梳理了 MiniCPM 在老挝、马来西亚、肯尼亚等实际落地情况,读者可借此看小模型如何嵌入日常系统而非比拼参数量。
Exactly! 💙 The real impact of small models comes from making AI accessible, affordable, and useful in everyday life. Excited to see more real-world applications built with MiniCPM! 🚀
在 X 查看被引用的帖子MiniCPM just became a country's official model — and it only has 8 billion parameters.
ModelBest says MiniCPM downloads passed 70 million as of November 2026. Its 8B Laos national model is now plugged into the government's G-Chat v2 messenger, and its Lao translation beat GPT-4o on an international benchmark.
Laos isn't the only deployment. Malaysia is using VoxCPM2 for voice interaction, a bank in Kenya runs on-device risk agents on it, and developers are shipping offline apps on top of it.
The race for small on-device models isn't about parameter counts anymore. It's about who actually gets embedded in daily systems — cheap enough and offline-capable enough to get in.
If you build on-device models or ship AI into the real world, will the next national-level deal go to an open model or a big lab's in-house one?
#MiniCPM #OnDeviceAI
来源:OpenBMB · x.com