Nous Research has released NousCoder-14B, an open-source competitive programming model that it says matches or beats several larger proprietary systems, trained in just four days on 48 of Nvidia's latest B200 graphics processors. The release lands as Anthropic's Claude Code dominates developer conversation, underscoring how fiercely companies large and small are now competing in AI-assisted software development.

NousCoder-14B reaches a 67.87 percent accuracy rate on LiveCodeBench v6, a standardized set of competitive programming problems published between August 2024 and May 2025. That is a 7.08 percentage-point gain over its base model, Alibaba's Qwen3-14B, a sizable jump for a model this compact in the open-source AI model field.

Radical openness sets it apart

What truly distinguishes the release is its transparency. Nous Research published not only the model weights but the complete reinforcement learning environment, benchmark suite, and training harness built on its Atropos framework. That means any researcher with enough compute can reproduce or extend the work, a level of openness rare in a field dominated by closed systems. The model is available now on Hugging Face under a permissive Apache 2.0 license.

The project was led by Joe Li, a researcher in residence and former competitive programmer. In his technical report, Li compared the model's leap, from roughly a 1600 to 1750 rating range up to 2100 to 2200, to his own two-year climb on Codeforces between ages 14 and 16. The model covered the equivalent in four days, though he noted he solved about 1,000 problems in that span while the model needed 24,000, a reminder that humans remain far more sample-efficient learners.

Inside the training system

NousCoder-14B is trained on "verifiable rewards": the model writes code, the solution runs against test cases, and it gets a binary correct-or-incorrect signal. Nous Research used the cloud platform Modal to run sandboxed code execution in parallel, with each problem averaging hundreds of test cases under 15-second and 4-gigabyte limits. Key techniques included:

  • DAPO, or Dynamic Sampling Policy Optimization, which discards examples the model always solves or always fails
  • Iterative context extension, starting at 32,000 tokens, expanding to 40,000, and reaching best results near 80,000 during evaluation
  • Pipelining that overlaps inference and verification to maximize GPU utilization

A looming data shortage

Li's report carries a finding with broad implications: the training set already covers much of the readily available, verifiable competitive programming material on the internet. In other words, this domain is approaching the ceiling of high-quality data. He argued that future progress will lean heavily on synthetic data generation and more data-efficient algorithms, and pointed to training models to generate their own solvable problems as a promising route toward self-play.

A bet on open-source AI

Nous Research, backed by crypto venture firm Paradigm, has raised $65 million in total and positions its open-source releases as direct competitors to proprietary alternatives. Previous work includes the Hermes 4 family and DeepHermes-3. The new model drew praise for advancing reproducible research and skepticism about benchmark-focused marketing, a tension common to fast-moving AI launches. The deeper question it raises is no longer whether machines can learn to code, but whether they will soon write their own training problems and teach themselves. Track the field in our latest tech news.

Frequently asked questions

What is NousCoder-14B?

It is an open-source, 14-billion-parameter coding model from Nous Research, built on Alibaba's Qwen3-14B and tuned for competitive programming. It scores 67.87 percent on LiveCodeBench v6 and is released under the Apache 2.0 license.

How was it trained so quickly?

Nous Research used reinforcement learning with verifiable rewards on 48 Nvidia B200 GPUs over four days, running sandboxed code tests in parallel through the Modal platform and pipelining inference with verification.

Can I run or reproduce NousCoder-14B?

Yes. The weights are on Hugging Face, and Nous Research also published the full RL environment, benchmark suite, and training harness, so researchers with enough compute can reproduce or build on the results.