Nous Research's NousCoder-14B is an open-source coding model in the Claude Code era
Nous Research released NousCoder-14B, an open-source coding model achieving 67.87% Pass@1 on LiveCodeBench v6 (+7.08% over base Qwen3-14B). Trained with DAPO (Dynamic Sampling Policy Optimization) on 48 Nvidia B200 GPUs over 4 days using 24,000 competitive programming problems. Fully open source under Apache 2.0 license, competing in the Claude Code and Codex CLI era.

Nous Research's NousCoder-14B: An Open-Source Coding Model Built for the Claude Code Era

Nous Research has unveiled NousCoder-14B, an open-source coding model that enters the rapidly evolving AI coding assistant space at a pivotal moment. The 14-billion parameter model demonstrates that open-source alternatives can compete with proprietary coding agents like Claude Code and Codex CLI, achieving 67.87% Pass@1 on LiveCodeBench v6 — a 7.08% improvement over its base model Qwen3-14B.
Technical Innovation: DAPO Training

The model was trained using DAPO (Dynamic Sampling Policy Optimization) on 48 Nvidia B200 GPUs over four days with 24,000 competitive programming problems. Fully open source under Apache 2.0 license on Hugging Face, with the Atropos training stack on GitHub. Joe Li noted that competitive programming data is approaching its limits — future gains will come from synthetic data and self-play.

