Apparently, you can build tiny reasoning models via LoRA.
Tina is a family of 1.5B models trained using LoRA-based RL—cost-efficient and performant.
The best Tina hits >20% gain in reasoning performance and 43% Pass@1 on AIME24,
with just $9 in post-training + eval cost.
Tiny Reasoning Models via LoRA Achieve Strong AIME Performance
By
–
