Skip to content
2 modelspublished as public checkpoints
3Bparameter scale, small enough to self-host
Structuredreasoning before the final answer

The problem

Small models often answer directly instead of reasoning through a problem, which costs accuracy on multi-step maths and logic.

Approach

  • Applied Group Relative Policy Optimization, the approach used in DeepSeek-R1, as post-training on 3B models.
  • Trained the models to generate structured reasoning before producing a final answer.
  • Published both Llama 3.2 3B and Qwen 2.5 3B variants on Hugging Face.

Stack

  • PyTorch
  • Hugging Face
  • GRPO
  • TRL
  • Llama 3.2
  • Qwen 2.5

What these numbers do and do not prove

These are research checkpoints for experimentation, not production-hardened models. Reproduce the training and evaluation before relying on the results.

Source

Other work

Start a conversation