Tencent Upgrades Self-Developed Foundation Model Tencent Hunyuan and Deploys it to Internal Products


Tencent Hunyuan announced on November 18, 2026, that it is open-sourcing the Hy4flash model, focusing on extreme price-performance ratio. This open-source release includes three versions—Hy4flash, Hy4flash-FP8, and Hy4flash Base—forming a family matrix and supporting convenient deployment across multiple channels. To help global developers get started quickly and conduct secondary development, Tencent has simultaneously opened multiple mainstream code channels, boosting the AI open-source ecosystem.
Tencent Hunyuan has released a brand-new real-time voice model, Hy Realtime. Focusing on real-time performance, it excels in basic conversational capabilities such as speech understanding, instruction following, and human-like expression, and newly adds Agentic tool calling and singing capabilities. It upgrades from simple responses to executable tasks and is closer to real human conversation. The model has been applied to products such as Yuanbao voice calls and Tencent Customer Service.
Three months ago, Tencent Hunyuan launched Hyra 1.0, a minimalist general-purpose Agent loop system for scientific research. Its open-source model Hy3 drives Hyra to solve a core problem in additive combinatorics that had remained unresolved for over fifty years—the optimal exponents for sum sets and difference sets—setting a record as the first Chinese open-source model to fully automatically solve an open problem in mathematics. Since then, international AI giants have been competing in mathematics and theoretical computer science. The "AI problem-solving" path has been proven viable, and the next step of automated scientific research is drawing attention.
Tencent Hunyuan launches 'Tencent Hy Translate' app, available on mini-program, plugin and PC. Powered by Hy-MT2, it supports 33 languages, including 5 Chinese minority languages/dialects, with voice, photo and offline translation. Offline models can be downloaded, covering HK, Macau, Taiwan and Southeast Asia.....
Tencent Hunyuan research examines batch size in online LLM RL. As models self-generate data and rollout/training scaling rates diverge, classical critical batch size theory needs revisiting. Covering GRPO and PPO, it gives practical guidance for efficient training on large GPU clusters.....