Tencent Releases Its Open-Source Hy3: Small Model, Big Claim

Tencent has released its Hy3 AI model as open source — at a price that's easy on the wallet. By the company's account, Hy3 rivals flagship open-source models 2-5x its size despite being small, after post-training was strengthened with feedback from 50+ products following an April preview.

In brief

• What launched: Hy3, Tencent's new large language model, open-sourced under Apache 2.0 on GitHub, Hugging Face, and ModelScope.
• The claim: per Tencent, Hy3 rivals flagship models 2-5x its size. In a blind evaluation with 270 experts using tasks from their own work, it scored 2.67/4, beating GLM-5.1 (2.51); it's strongest on frontend, data, and CI/CD.
• Concrete gains: in internal tests, hallucination dropped from 12.5% to 5.4% and multi-turn issues from 17.4% to 7.9%. In one internal product (WorkBuddy), task success rose from 72% to 90% and completion time fell 34%.
• Price: about 1 RMB input / 4 RMB output per 1M tokens — very cheap — and being open source, you can also run it on your own infrastructure.

Our take

The value for you is clear: there's another cheap, open option on the table. Hy3's pitch isn't to be the smartest model — it's that, despite being small and cheap, it handles everyday work (code, office tasks, slides, frontend) well. One figure Tencent shared stands out: on a comparable task it used nearly half the tokens of GLM-5.2 — the same work for less money. These numbers are largely Tencent's own tests, so don't take them on faith — but if you have templated work (reports, slides, simple code), it's a cheap candidate worth trying. And because it's open source, you can run it in your own environment without sending your data out.

Kaynak: Tencent Hunyuan