NousCoder-14B

📰 Article[email protected] (Michael Nuñez)

Open-source coding model matches larger proprietary systems, trained in 4 days on 48 B200 GPUs.

NousCoder-14B is a 14-billion parameter open-source coding model from Nous Research, trained in just four days using 48 Nvidia B200 GPUs. It achieves 67.87% accuracy on LiveCodeBench v6, a 7.08% improvement over its base model Qwen3-14B. The training employed DAPO (Dynamic Sampling Policy Optimization) with verifiable rewards: the model generates code, executes it against test cases, and receives a binary correct/incorrect signal. Dynamic sampling discards examples where the model solves all or fails all attempts to ensure useful gradient signals. Iterative context extension expanded from 32K to 40K tokens during training, with evaluation up to 80K tokens yielding best results. The pipeline overlaps inference and verification to maximize GPU utilization. All resources—model weights, Atropos framework, benchmark suite, and training harness—are open-sourced, enabling full reproducibility. Researcher Joe Li, a former competitive programmer, noted the model's improvement from ~1600-1750 to 2100-2200 Codeforces rating equivalent in four days, a leap that took him two years as a human, though the model required 24,000 problems versus his 1,000.

NousCoder-14B

💡Highlights

🎯For

🔗Links