Tencent Hunyuan releases UniRL: unified multimodal reinforcement learning infrastructure
Original title:🚀Introducing UniRL, an RL infra for unified multimodal models. Together with two new RL algorithms:...
AI summary
Tencent Hunyuan launched UniRL, a reinforcement learning infrastructure supporting unified multimodal models, and released two new algorithms, DRPO and Flow-DPPO. UniRL covers diffusion/flow-matching models, LLM/VLM, and unified multimodal models (such as Hunyuan-Image 3 and Bagel) through a single post-training loop (generation → scoring → advantage → update → synchronization). Models and algorithms serve as independent axes, enabling combined coverage of model × algorithm. The framework supports pluggable rollout engines (training-side/SGLang/vLLM-Omni), FSDP2 sharding, and three deployment modes. FlowDPPO introduces trust-region policy optimization based on exact divergence for flow/diffusion models; DRPO provides a smooth advantage-weighted quadratic regularization method for LLM RL. The code is open-sourced.
Key Takeaways
- Tencent Hunyuan launched UniRL,a reinforcement learning infrastructure supporting unified…
- UniRL covers diffusion/flow-matching models,LLM/VLM,and unified multimodal models (such as…
- Models and algorithms serve as independent axes,enabling combined coverage of model × algorithm.
- The framework supports pluggable rollout engines (training-side/SGLang/vLLM-Omni),FSDP2…
- FlowDPPO introduces trust-region policy optimization based on exact divergence for flow/diffusion…
Quick Facts
- Published
- Updated
- Source
- X:腾讯混元 (@TencentHunyuan)
- Category
- Tutorials & Tips
- AI score
- 67
UniRL把扩散和LLM的强化学习塞进同一个训练循环,外加两个新算法,多模态对齐的研究者可以立刻fork代码试起来。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Tencent Hunyuan releases UniRL: unified multimodal reinforcement learning infrastructure. June 9, 2026. https://new.nat6.net/en/news/cmq6kwvh809eesl5irgb0uzf2/