Nous Research, a startup backed by crypto venture firm Paradigm, has released a new competitive programming model that matches or exceeds larger proprietary systems. The model, called NousCoder-14B, was trained in just four days using 48 Nvidia B200 graphics processors and achieves a 67.87 percent accuracy rate on LiveCodeBench v6.
The release is significant as it marks another entry in the crowded field of AI coding assistants. However, Anthropic's Claude Code has dominated social media discussion with its demonstrations of end-to-end software development capabilities. Nous Research is betting that open-source alternatives trained on verifiable problems can close the gap and transparency in how these models are built matters.
The model was built using a reinforcement learning system that trains on 24,000 competitive programming problems. The approach relies on "verifiable rewards" where the model generates code solutions, those solutions are executed against test cases, and the model receives a binary signal: correct or incorrect. This feedback loop requires significant infrastructure to execute at scale.
The researchers also highlighted the looming data shortage that could slow AI coding model progress. They found that they had approached the limits of high-quality training data within the competitive programming domain, suggesting that future research will focus on synthetic data generation and data-efficient algorithms and architectures.
