AReaL v0.2 releases 7B and 32B reasoning models with their training data
- Source
- x.com
- Date
We're thrilled to unveil AReaL v0.2 boba! #LRM #AI ✅ SOTA 7B model: 61.9 AIME24 & 48.3 AIME25 ✅ 32B model: match QwQ on AIME24 with only 200 samples! ✅ Full #SGLang support Training data/weights ALL OPENSOURCE🔥🔥🔥 Like A ReaL boba, for everyone! 🧋https://t.co/OopXbE7sTJ https://t.co/rEn9g7sJLS
Context
Training a that beats a strong existing baseline usually means throwing a large amount of compute and data at reinforcement learning. Ant Open Source's AReaL v0.2 release reports a 7B model scoring 61.9 on AIME24 and 48.3 on AIME25, and says a 32B model matches Alibaba's QwQ on AIME24 using only 200 training samples. The team published the training data and model weights alongside the announcement, so the sample-efficiency claim is something a reader can attempt to reproduce rather than only read about.
- reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
Checking sign-in…
Loading comments…