Nous Research Launches Open-Source NousCoder-14B AI Coding Model

Nous Research, an open-source artificial intelligence startup backed by crypto venture firm Paradigm, announced the release of its new competitive programming model, NousCoder-14B, on Monday. The model was trained in four days using 48 of Nvidia's B200 graphics processors and is reported to match or exceed the performance of several larger proprietary systems. This launch occurs amid increasing activity in AI coding assistants, notably the recent discussions around Anthropic's Claude Code.

NousCoder-14B achieved a 67.87% accuracy rate on LiveCodeBench v6, a standardized evaluation for competitive programming problems. This performance represents a 7.08 percentage point improvement over the Alibaba Qwen3-14B model, from which NousCoder-14B was trained, according to Nous Research's technical report.

A distinctive feature of the NousCoder-14B release is its open-source nature. Nous Research has published not only the model weights but also the complete reinforcement learning environment, benchmark suite, and training harness, built on its Atropos framework. This approach aims to enable researchers with adequate computing resources to reproduce or extend the work.

The model's training process involved a reinforcement learning system that utilizes "verifiable rewards." This method generates code solutions, executes them against test cases, and provides binary feedback (correct or incorrect). Nous Research employed Modal, a cloud computing platform, to run sandboxed code execution in parallel across 24,000 training problems. The training also incorporated techniques such as Dynamic Sampling Policy Optimization (DAPO) and iterative context extension, with inference and verification pipelined for optimized hardware utilization.

The technical report also highlighted a potential challenge regarding data scarcity. The 24,000 problems used for NousCoder-14B's training encompass a significant portion of readily available, verifiable competitive programming problems. This suggests that future advancements in this domain may require a focus on synthetic data generation and data-efficient algorithms, possibly through models that can generate solvable problems for self-play.

(Source: VentureBeat AI)