Welcome to RLinf#
Welcome to RLinf
Scalable RL Post-Training for Foundation Models and Embodied Agents
RLinf is a flexible, scalable open-source infrastructure for post-training foundation models with reinforcement learning. The “inf” stands for Infrastructure — a robust backbone for next-generation training — and for Infinite, capturing open-ended learning and continuous generalization.
Choose Your Path#
Fine-tune a VLA on LIBERO, ManiSkill, RoboTwin, and more with PPO or GRPO.
Browse agentic and reasoning recipes for Qwen / DeepSeek models.
Add a model, environment, or algorithm and plug it into RLinf.
Collocated, disaggregated, and hybrid placement across GPUs and nodes.
Why RLinf#
Strength |
What it gives you |
|---|---|
Fast |
Hybrid fine-grained pipelining delivers 120%+ throughput over comparable frameworks, plus automatic online scaling. |
Flexible |
Switch FSDP + Hugging Face for prototyping or Megatron + SGLang for large-scale training, with no code changes. |
Proven |
Built-in PPO, GRPO, DAPO, and Reinforce++, with SOTA recipes for embodied and reasoning tasks. |