From the source
Ant Ling released AReaL v0.2 (Boba), an open-source reinforcement learning framework.
The release includes a SOTA 7B model for mathematical reasoning (pass@1 61.9 on AIME24, 48.3 on AIME25) and a competitive 32B model trained with only 200 data samples that matches QwQ-32B on AIME 2024.
The framework adds SGLang support and achieves 1.5x speed improvement over v0.1 on 7B models.
