Tencent-Hunyuan/UniRL v0.1.0

UniRL v0.1.0

Repository: Tencent-Hunyuan/UniRL

Tag: v0.1.0

Published: 2026-10-07T03:47:51Z

Prerelease: no

Release notes: Initial public release of UniRL, a reinforcement learning framework for unified multimodal models. Install with pip install unirl==0.1.0.

What's included

• Training entrypoints: diffusion RL (train_diffusion), AR RL for LLM / VLM (train_ar), prompt enhancement (train_pe), and unified AR + diffusion RL (train_unified_model).
• Algorithms: Flow-DPPO and DRPO (team-proposed), plus GRPO, FlowGRPO, and DiffusionNFT.
• Models: Stable Diffusion 3.5, Qwen-Image, FLUX.2-Klein, WAN 2.1, WAN 2.2, HunyuanVideo 1.0 / 1.5, HunyuanImage 3.0, BAGEL, Qwen3, Qwen2.5-VL, and the Prompt-Enhancer composition.
• Rollout engines: trainside, SGLang (diffusion and LLM), and vLLM-Omni, on a Ray + FSDP runtime with LoRA / full-weight sync.