Tencent Hunyuan Launches Hy3 Preview: Open‑Source Model Boosts Agent Performance

On April 23, Tencent announced the open‑source release of Hy3 preview, the first model rebuilt after the Hunyuan reconstruction. The hybrid expert model contains 295 B total parameters, 21 B active parameters, and supports a maximum context length of 256 K tokens.

The redesign follows three practical principles established in February 2026: (1) Systematic capability integration to avoid “specialization” and ensure deep cooperation among reasoning, long‑text handling, instruction following, dialogue, code, and tool use; (2) Authenticity evaluation that avoids leaderboard‑driven bias by employing self‑built tests, latest exams, manual reviews, and product crowd‑testing; (3) Cost‑performance focus, where a tightly coupled model architecture and inference framework dramatically reduce task costs, making intelligent services affordable and effective.

Hy3 preview markedly improves context learning and instruction‑following abilities. Inspired by Tencent’s business scenarios, the team introduced CL‑bench and CL‑bench‑Life to assess these skills, and the model shows significant gains on these benchmarks.

Complex reasoning, a foundation for solving diverse problems, is demonstrated by strong results on FrontierScience‑Olympiad, IMOAnswerBench, and the 2026 Tsinghua University Mathematics Qualification Exam (Spring 26) as well as the national high‑school biology competition (CHSBO 2025), indicating robust general‑purpose reasoning.

Code generation and agent capabilities see the most pronounced improvements. Thanks to the rebuilt pre‑training and reinforcement‑learning pipeline, Hy3 preview achieves competitive scores on SWE‑Bench Verified, Terminal‑Bench 2.0, BrowseComp, WideSearch, ClawEval, and WildClawBench.

Internal evaluations on Hy‑Backend, Hy‑Vibe Bench, and Hy‑SWE Max further confirm the model’s competitiveness across backend engineering tasks and high‑difficulty software‑engineering benchmarks.

Inference efficiency is boosted by 40 % through deep co‑design of the model and inference framework, including operator optimizations and quantization algorithms, resulting in a substantial cost reduction compared with the previous generation. On Tencent Cloud’s TokenHub, input pricing is as low as ¥1.2 per million tokens (cache hit ¥0.4) and output pricing starts at ¥4 per million tokens.

Hy3 preview has been rolled out to core Tencent AI products such as Yuanbao (co‑designed for intent understanding, text creation, and deep search), ima knowledge‑base QA, CodeBuddy, WorkBuddy, AI NPC in Peace Elite, QQ AI Assistant, and Tencent Docs PPT generation. Notable real‑world gains include a 54 % reduction in first‑token latency, a 47 % drop in end‑to‑end latency, and success rates exceeding 99.99 % on CodeBuddy/WorkBuddy; PPT generation success rose 20 %, evaluation scores improved 10 %, and generation time fell 20 %.

The team invites feedback from the open‑source community and users to further enhance the upcoming official Hy3 release, while continuing to scale pre‑training and reinforcement‑learning to raise the model’s intelligence ceiling.