ByteDance and Tsinghua open-source DAPO, an RL system hitting 50 on AIME 2024

ByteDance's Seed team and Tsinghua University's AIR lab have released DAPO, a complete open-source system for reinforcement learning on large language models. According to the project's GitHub repository, which surfaced on the Hacker News front page, the released checkpoint — DAPO-Qwen-32B, trained from the Qwen2.5-32B base model — scores 50 points on the AIME 2024 math benchmark, which the team says surpasses the previous best from DeepSeek-R1-Zero-Qwen-32B while using half the training steps.

What is being released

DAPO stands for Decoupled Clip and Dynamic sAmpling Policy Optimization. The release goes beyond a paper: it covers the algorithm, the training and validation datasets, the training scripts, the model weights, and the infrastructure code needed to rerun the pipeline. Everything is built on the open-source verl RL framework, and the team notes the experiments were run on the Volcano Engine Machine Learning Platform, with a fuller reproduction guide for that platform promised later.

The benchmark numbers

The headline result is 50 points on AIME 2024 starting from a Qwen2.5-32B base model. The repository describes this as state-of-the-art performance for large-scale LLM reinforcement learning, overtaking DeepSeek-R1-Zero-Qwen-32B with 50% of the training steps.

The release arrived in stages. A March 2025 update published the training record for an early variant that lacked two components — token-level policy gradient loss and dynamic sampling — which reached 44 points on AIME 2024. A May 2025 update then added the training logs for the full DAPO recipe together with the checkpoint that crosses 50 points, plus instructions for evaluating the model on AIME 2024.

What the training curves show

The maintainers publish their metrics through Weights & Biases, and the README highlights three trends they consider signs of healthy training. Response length grows steadily over the run, which they argue enables greater exploration and helps the model learn more complex reasoning behaviour. The reward signal rises in a stable fashion without large fluctuations, suggesting the model is fitting the training distribution consistently. And entropy first drops, then climbs in a controlled way, which the team says keeps exploration and exploitation in balance and steers clear of both overfitting and excessive randomness.

Reproducing the run

Training uses DAPO-Math-17k, a curated and processed mathematics dataset, with AIME 2024 serving as the validation set. The repository ships out-of-the-box scripts for both configurations, and the maintainers state that each has been verified against the current verl release: the ablated variant reproduces the 44-point score, and the full pipeline reproduces the 50-point score, with the corresponding training records viewable online.

For inference, the README provides example code using vLLM with bfloat16 precision and a tensor-parallel setup across eight GPUs. Benchmark evaluation on AIME 2024 is handled through Ray Serve and vLLM, and the model weights are distributed via Hugging Face under the identifier BytedTsinghua-SIA/DAPO-Qwen-32B.

Why it matters

Reasoning ability in modern LLMs is largely forged through reinforcement learning after pretraining, and the exact recipes used by frontier labs are rarely disclosed. DAPO releases the entire chain — algorithm, data, code, weights, and training logs — which turns a benchmark claim into something the wider community can verify, audit, and build on. The result also demonstrates that a mid-sized 32B base model can reach competitive math performance via an open method rather than a proprietary pipeline. For research groups without frontier-lab resources, the verified scripts, the published dataset, and the transparent training records meaningfully lower the barrier to experimenting with RL-trained reasoning models.