Nous Research's NousCoder-14B is an open-source coding model landing right in the Claude Code moment

Overview

Nous Research, an open-source AI startup backed by Paradigm, released NousCoder-14B, a competitive programming model it says matches or exceeds several larger proprietary systems. The model was trained in four days using 48 of Nvidia's B200 GPUs and arrives as Anthropic's Claude Code dominates discussion among developers. Nous Research published the full model weights, reinforcement learning environment, benchmark suite and training toolkit to make the work reproducible.

Key Takeaways

NousCoder-14B is an open-source competitive programming model from Nous Research, backed by Paradigm.
It was trained in just four days using 48 of Nvidia's latest B200 GPUs.
The model scores 67.87 percent accuracy on LiveCodeBench v6, a 7.08 percentage point gain over its base model, Alibaba's Qwen3-14B.
Nous Research published the model weights, reinforcement learning environment, benchmark suite and training toolkit, built on its Atropos framework.
The release arrives as Anthropic's Claude Code dominates developer discussion since New Year's Day.
The model required 24,000 problems, while its trainer Joe Li solved roughly 1,000 over two years, underscoring that humans remain more sample-efficient.

Stats & Key Facts

#48 Nvidia B200 GPUs used for training
#Trained in four days
#67.87 percent accuracy on LiveCodeBench v6
#7.08 percentage point improvement over Qwen3-14B
#Model trained on 24,000 competitive programming problems
#Li solved roughly 1,000 problems over two years

Nous Research's NousCoder-14B is an open-source coding model landing right in the Claude Code moment

What Nous Research released

The model targets competitive programming.

›NousCoder-14B is a new competitive programming model released on a Monday.
›Nous Research says it matches or exceeds several larger proprietary systems.
›It was trained in four days using 48 of Nvidia's latest B200 GPUs.

Nous Research is an open-source AI startup backed by crypto venture firm Paradigm. The model enters a crowded field of AI coding assistants.

The Claude Code moment

›Claude Code is Anthropic's agentic programming tool.
›It has dominated social media discussion since New Year's Day.
›Developers have posted testimonials about its capabilities.

Jaana Dogan, a principal engineer at Google responsible for the Gemini API, wrote in a viral post that Claude Code generated in an hour what her team built last year. She was describing a distributed agent orchestration system her team had spent a year developing, which Claude Code approximated from a three-paragraph prompt.

Benchmark performance

›NousCoder-14B achieves 67.87 percent accuracy on LiveCodeBench v6.
›That is a 7.08 percentage point improvement over the base model, Alibaba's Qwen3-14B.
›LiveCodeBench v6 tests competitive programming problems published between August 2024 and May 2025.

The benchmark figures come from Nous Research's technical report published alongside the release.

Radical openness

›Nous Research published the complete model weights.
›It also released the reinforcement learning environment, benchmark suite and training toolkit.
›The stack is built on the company's Atropos framework.

The release enables any researcher with sufficient compute to reproduce or extend the work. One observer on X said open-sourcing the Atropos stack provides the necessary infrastructure for reproducible olympiad-level reasoning research.

The trainer's story

›The model was trained by Joe Li, a researcher in residence at Nous Research and a former competitive programmer.
›Li compared the model's improvement to his own journey on Codeforces.
›He estimated the model moved from roughly the 1600-1750 rating range to 2100-2200.

Li said that leap mirrors one that took him nearly two years of sustained practice between ages 14 and 16, while the model accomplished the equivalent in four days. He called watching the final training run a surreal experience.

Sample efficiency caveat

›Li solved roughly 1,000 problems during those two years.
›The model required 24,000 problems.
›Humans, at least for now, remain dramatically more sample-efficient learners.

Li noted this caveat as a point about broader questions of AI efficiency, contrasting the model's data needs with human learning.

Frequently Asked Questions

What is NousCoder-14B?

It is an open-source competitive programming model from Nous Research that the company says matches or exceeds several larger proprietary systems.

How was it trained?

It was trained in four days using 48 of Nvidia's latest B200 GPUs, on 24,000 competitive programming problems, using the Atropos framework.

How does it score on benchmarks?

It achieves 67.87 percent accuracy on LiveCodeBench v6, a 7.08 percentage point improvement over its base model, Alibaba's Qwen3-14B.

What did Nous Research open-source?

It published the model weights, reinforcement learning environment, benchmark suite and training toolkit, enabling reproduction or extension of the work.

What does the sample-efficiency comparison show?

The model required 24,000 problems, while trainer Joe Li solved roughly 1,000 over two years, showing humans remain dramatically more sample-efficient.

Nous Research positions NousCoder-14B as a fully reproducible open-source coding model arriving amid the Claude Code surge.