BroadLink says Huansi Hub cuts AI appliance cost to about RMB1 per device per year
BroadLink founder and CEO Liu Zongru told Geek Park's Hardwire that the company's Huansi Hub is built from a set of very small expert models — each possibly under 10B parameters — with a large model moved to the background as a "grader." He said the system-1 layer only judges and executes rather than generating text, with a claimed hit rate above 95% and a 99% target, and that DeepSeek's latest model grades the previous day's interaction logs each night and distills its judgments back into the small models. Liu said the hub has already been deployed with B2B customers across several appliance categories, that AI cost can be compressed to roughly RMB1 per device per year, and that existing Wi-Fi plus Bluetooth appliances need only a module OTA upgrade so users can talk to them through a WeChat mini-program assistant; all figures are the company's own claims.
The interview offers a detailed, vendor-attributed example of a cost-focused architecture for putting AI into mass-market appliances: tiny expert models handle most commands while a frontier model is used mainly for nightly retraining.
过去两年多，在智能硬件领域，投资人和创业者都热衷于为 AI 寻找一个个物理世界里的新「容器」，各种新形态的 AI 原生硬件成了投资市场的热门项目。
但其实，很多成熟形态的硬件同样有 AI 化的需求。这也许没有「AI Native」的概念那么性感，但市场一点不小，复杂度丝毫不低。
以家电行业为例：供应链长，改一次硬件就要重走研发、测试、认证；销售链条长，从厂家、经销商到上门安装的电工，每一环都得接得住；成本又极度敏感，多一块钱都要反复掂量。
好在，家电已经为 AI 打好了地基。过去十几年，大量家电已经具备联网能力——这意味着灵活升级的 OTA「管道」已经铺好，行业缺的是让设备真正被用起来的解决方案。
那么，硬件到底需要什么样的 AI？9 月，TypeSafe 推出的 Jev 在开发者圈引发了不少讨论，它提供了一个参照：Jev 不生成文字，只输出带置信度的判断。在硬件乃至具身领域的不少场景里，判断比生成更重要。