DAPO, short for Decoupled Clip and Dynamic Sampling Policy Optimization, is an openly released system for training large language models to reason through reinforcement learning. A GitHub repository maintained by BytedTsinghua-SIA, a joint effort between ByteDance Seed and Tsinghua University’s AIR institute, hosts the code, training data, and model weights. The changelog dates the project to March 2025, not 2026, when the team posted training records for an early version. A May 2025 update added a checkpoint and evaluation instructions. Nothing in the repository points to a 2026 release; this is a year-old project still getting cited as new.
The team built DAPO on Qwen2.5-32B, an open base model. According to the repository, the system scored 50 points on the AIME 2024 math benchmark, which the maintainers say surpassed DeepSeek’s comparable R1-Zero-Qwen-32B model using half the training steps. That comparison, like the benchmark result itself, is the team’s own reported claim, not an independently verified score. The project builds on verl, an existing open-source RL framework the maintainers credit for the underlying infrastructure.
For engineers evaluating open RL pipelines now, test DAPO against 2026-era alternatives before assuming it still represents current best practice.
Details per the DAPO repository on GitHub, maintained by BytedTsinghua-SIA (ByteDance Seed and Tsinghua AIR), last updated May 2025.
