coding

NousResearch Boosts Qwen3-14B Coding to 68% Pass@1

NousResearch enhances Qwen3-14B's coding performance to achieve 68% pass@1 rate through advanced fine-tuning techniques and optimization strategies for

Someone at NousResearch took Qwen3-14B and turned it into a pretty solid coding model through reinforcement learning.

The results from LiveCodeBench v6:

  • Base Qwen3-14B: 60.79% Pass@1
  • NousCoder-14B: 67.87% Pass@1
  • That’s a 7.08% jump in accuracy

They trained it on 24k coding problems using 48 B200 GPUs over four days. The model is available at https://huggingface.co/NousResearch/NousCoder-14B

What’s interesting is they focused specifically on competitive programming problems - the kind where there’s a clear right/wrong answer. This makes RL training way more practical since the model gets immediate feedback on whether its code actually works.

For anyone doing coding tasks that need reliable solutions rather than just “sounds about right” responses, this kind of specialized training seems to make a real difference.