Nous Research Launches NousCoder-14B, an Open-Source Model Matching Proprietary Coding Assistants

Nous Research, an open-source AI startup backed by crypto-focused venture firm Paradigm, unveiled NousCoder-14B, a competitive programming model that claims to match or surpass several larger proprietary systems. The model was trained in just four days using 48 of Nvidia's latest B200 graphics processors, showcasing an efficient approach to AI model development. The release arrives at a particularly significant moment for the industry: Claude Code from Anthropic has dominated social media discussions since early January, with developers enthusiastically praising its capabilities. This timing underscores the fierce competition among companies seeking to establish dominance in AI-assisted software development, a space many believe will become foundational to how software is built. In standardized testing, NousCoder-14B achieved a 67.87% accuracy rate on LiveCodeBench v6, representing a 7.08 percentage point improvement over its base model, Alibaba's Qwen3-14B. What particularly distinguishes this release is its radical transparency: alongside the model weights, Nous Research published the complete reinforcement learning environment, benchmark suite, and training infrastructure—built on the company's Atropos framework—enabling any researcher with sufficient computational resources to reproduce or extend the work. The training employed a sophisticated reinforcement learning approach based on "verifiable rewards," where the model generates code solutions that are executed against test cases, receiving binary feedback. The system trained on 24,000 competitive programming problems, each containing hundreds of test cases on average. The approach utilized DAPO (Dynamic Sampling Policy Optimization) and "iterative context extension," beginning with a 32,000-token context window before expanding to 40,000 tokens, and ultimately evaluating at approximately 80,000 tokens. A crucial finding from the technical report reveals that researchers are approaching the limits of high-quality training data within competitive programming. The total number of available competitive programming problems on the internet is approximately the same order of magnitude as the 24,000 used for training, suggesting the domain is reaching data saturation. This observation reflects growing concerns across the AI industry about data constraints, pointing toward future research focused on synthetic data generation and data-efficient algorithms. Nous Research, which previously released Hermes 4 and DeepHermes-3, solidified its position as a company committed to open-source releases competing with proprietary alternatives. The company secured $50 million in funding in April 2025, led by Paradigm, bringing total funding to $65 million. The model is available on Hugging Face under an Apache 2.0 license, with the complete Atropos training stack also published.