verl-project
verl-project is an open-source research and engineering project whose principal software, verl (also called HybridFlow), provides a flexible, efficient and production-ready framework for post-training large language models with reinforcement learning. It aims to make complex RL dataflows easier to express, execute and scale across heterogeneous hardware and existing LLM infrastructure.
At a Glance
- AI research teams and academic researchers
- LLM foundation-model developers and post-training teams
- Enterprises building reasoning models and agentic systems
- Cloud and GPU infrastructure users operating distributed training clusters
- +2 more
AI Tools by verl-project
(1)
verl
Open Source LLM RL Training Library
Discussions
No discussions yet
Be the first to start a discussion about verl-project
Latest News
Uni-Agent: Train Long-Horizon Agents at Scale
verl-vla v0.1.0: A Unified Cloud-Edge Post-Training Loop for Large-Scale VLA
Bringing End-to-End LLM Reinforcement Learning to AMD Instinct MI455X with verl
VeRL-Omni v0.2.0: Faster Diffusion RL and Stable Omni Training
Products & Services
Apache-2.0 open-source RL post-training framework for LLMs. It supports supervised fine-tuning, PPO, GRPO and many related algorithms, hybrid-controller execution, modular training/rollout backends, flexible device mapping, model-based or function-based rewards, multi-turn tool calling and scaling to very large models and GPU clusters.
Dedicated repository of end-to-end RL training recipes based on verl, including DAPO, SPPO, PRIME, GSPO, ReTool and other algorithm and application recipes.
Unified VLA post-training framework built on verl for human-in-the-loop data collection, supervised fine-tuning, evaluation and reinforcement learning across simulators, physical/teleoperation devices and distributed cloud-edge resources. v0.1.0 supports ACT, Gaussian Actor, Pi0.5 and Isaac GR00T, with LIBERO and Isaac Lab-Arena workflows.
RL stack for diffusion and omni-modal model post-training. v0.2.0 adds faster request-level-batched diffusion rollout, reusable omni adapters, Qwen3-Omni training, DPO/GSPO recipes, LTX-2.3 and Qwen-Image-Edit support.
Market Position
verl positions itself as a flexible, efficient and production-ready open-source RL post-training framework, differentiated by its hybrid-controller programming model, modular integration of training and inference engines, flexible device mapping and large-scale throughput. It overlaps with OpenRLHF, Hugging Face TRL, NVIDIA NeMo-Aligner and DeepSpeed-Chat; the HybridFlow paper says the project was inspired by Nemo-Aligner, DeepSpeed-Chat and OpenRLHF.
Founding Story
The project originated as HybridFlow, a research framework created to address the flexibility and efficiency problems of RLHF dataflows for large language models. The HybridFlow paper was submitted to arXiv on September 28, 2024; its source code was released as verl, initially under the Volcengine organization and later migrated to the verl-project GitHub organization in January 2026. The project is associated with ByteDance's Seed team and is developed as an open community project.
Business Model
Revenue Model
The project is distributed as open-source software under the Apache License 2.0 and provides public source code and documentation. The reviewed sources describe no paid plans, hosted API, subscription or license-fee model.
Target Markets
- AI research teams and academic researchers
- LLM foundation-model developers and post-training teams
- Enterprises building reasoning models and agentic systems
- Cloud and GPU infrastructure users operating distributed training clusters
- Robotics and vision-language-action research teams
- Developers using NVIDIA, AMD or Ascend accelerators
- Research and production post-training of LLMs
- Reasoning-model training for mathematics, coding and STEM tasks
- RLHF, PPO/GRPO and verifiable-reward training
- Multi-turn conversational agents and tool-using agents
- Large-scale distributed training of dense and MoE language models
- Vision-language, omni-modal and diffusion-model post-training
- ByteDance Seed / Doubao: the repository reports that Doubao-1.5-pro's RL scaling preview model was trained using verl.
- Mind Lab: the repository reports use of verl and Megatron-Bridge to train GRPO LoRA for a trillion-parameter model on 64 H800 GPUs.