AReaL
AReaL ProjectAReaL is a reinforcement learning system for training reasoning and agentic language models. It uses a fully asynchronous training paradigm that decouples rollout generation from policy updates, and supports FSDP and Megatron training backends with vLLM and SGLang inference backends. It was developed by the AReaL Team at Tsinghua IIIS and Ant Group.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public(areal-project/AReaL)
- core features withheld
- no — no enterprise or ee path in the repository root and no paid tier described in the README
Apache-2.0, an OSI license, with the full source public in this repository. There is no enterprise or ee path in the repository and no paid tier described in the README.
- https://raw.githubusercontent.com/areal-project/AReaL/main/LICENSE recorded 2026-10-07
LICENSE body, read in full: the stock Apache License 2.0 text, carrying no appended term or additional condition.
- https://raw.githubusercontent.com/areal-project/AReaL/main/README.md recorded 2026-10-07
README, read for the distribution and gating question: it installs and runs the code from this repository with "uv sync --extra cuda", which "installs training packages + SGLang (default inference backend)", and names no enterprise, pricing, paid or licence-gated build.
Adoption
1 medium confidenceAReaL's only PyPI release, 1.0.4, was uploaded 2026-06-10; the repository has since moved to the current v2.1.0 line, released 2026-08-25, so a download count from the old package measures a version the project has already left behind rather than current use. No stronger usage signal is available: GitHub stars reflect attention rather than use, so the level stays where it was.
- https://api.github.com/repos/areal-project/AReaL recorded 2026-08-31
Repo metadata for areal-project/AReaL: stargazers_count 5705, archived false, pushed_at 2026-08-31 - the attention figure the note contrasts with the download count.
- https://api.github.com/repos/areal-project/AReaL/releases/latest recorded 2026-08-31
tag_name v2.1.0, published_at 2026-08-25T10:23:59Z, prerelease false - the current stable release line, one major above the sole PyPI release.
- https://pypi.org/pypi/areal/json recorded 2026-08-31
info.version 1.0.4 with a single entry in `releases`, uploaded 2026-06-10T06:48:31Z. The package has made no other release, so the registry line has never moved past 1.x while the repository is on 2.1.0.
- https://pypistats.org/api/packages/areal/recent recorded 2026-08-31
last_month 45, last_week 9, last_day 3 for the areal package. Recorded as a measurement of the trailing registry line, explicitly NOT as the basis of the reading.
Capability
4 medium confidenceAReaL matches OpenRLHF's breadth: several RL algorithms running over FSDP or Megatron training with vLLM or SGLang rollout, scaled across tensor, pipeline and expert parallelism. It covers the same ground from a different angle, asynchronous agentic RL rather than OpenRLHF's Ray-scheduled RLHF at 70B and above, and neither reaches the wider method catalog of verl or nemo-rl.
- https://raw.githubusercontent.com/areal-project/AReaL/main/README.md recorded 2026-10-07
README describes "A Large-Scale Asynchronous Reinforcement Learning System" built on a fully asynchronous RL paradigm, a "Support Matrix" with Algorithms (GRPO, GSPO, PPO, DAPO, DPO, SFT among them), Models, "Training Backends" and "Inference Backends" tables, and example sets for Math & Reasoning, "Agentic RL" and "Vision-Language Models".
Verified 2026-10-07, except some scores, still awaiting re-confirmation