Tencent Yuanbao

A distinction worth drawing before anything else. Hunyuan is Tencent's model family and Yuanbao is the consumer application that serves it, and they are separate things with separate release cadences. Yuanbao is where a normal person in China meets Tencent's AI: an app, an iOS and Android download and a WeChat mini-program, sitting at or near the top five Chinese assistants by monthly active users alongside Doubao, Ernie Bot, Quark and Kimi. It got there partly on product and partly on a Spring Festival subsidy campaign reported at around one billion yuan, which is the sort of detail English coverage tends to leave out.

What makes it technically interesting is data access, not architecture. WeChat's official accounts and Channels hold an enormous volume of Chinese-language writing and video that sits behind a login wall, outside the index of every search engine including Baidu. Yuanbao can read it. For anyone doing market research, competitive tracking or technical reading in Chinese, that is a capability no amount of model quality substitutes for. On top of that the app carries a genuine model picker: Hunyuan and DeepSeek side by side since the DeepSeek-R1 integration in February 2025, Hunyuan Hy3 from 6 July 2026 at 295B total and 21B active parameters, and a Hy4 preview at 770B total with 49B active and a 1M token context window appearing in grey testing during August 2026, tagged as an expert tier above both. Agent mode, which produces downloadable PPT, Word, Excel, PDF and HTML files, went free for every user with the Hy3 launch.

The constraints are real and they are geographic. Registration wants a mainland phone number, the interface assumes Mandarin and there is no international build, so from outside China this is effectively closed. There is no Yuanbao API either. If you want to build on these models you go to Tencent Cloud TokenHub, where Hy3 input pricing dropped to 1 yuan per million tokens in July 2026, and full mainland cloud access generally means a Chinese business licence and real-name verification. Content moderation applies as it does across every domestic assistant: politically sensitive queries produce refusals or redirections, and if your evaluation touches policy, history or governance you need to account for that rather than discover it mid-project. Use it for what it is uniquely good at, which is reading the Chinese internet you otherwise cannot reach.