Nous Research, the open-source artificial intelligence startup backed by crypto venture firm Paradigm, released a new competitive programming model on Monday that it says matches or exceeds many large proprietary systems — trained in just four days using 48 of Nvidia’s latest B200 graphics processors.
The model, called NousCoder-14B, is another entry in the crowded field of AI coding assistants, but arrives at a particularly charged moment: Cloud Code, rival Anthropic’s agentic programming tool, has dominated social media discussion since New Year’s Day, with developers posting breathless testimonials about its capabilities. The simultaneous developments underscore how rapidly AI-assisted software development is evolving — and how companies large and small are competing to capture what many believe will become a foundational technology for how software is written.
Type: embedded-entry-inline Identification: 74cSyrq6OUrp9SEQ5zOUSl
NousCoder-14B achieves a 67.87 percent accuracy rate on LiveCodeBench v6, a standardized assessment that tests models on competitive programming problems published between August 2024 and May 2025. According to a technical report from Nous Research published alongside the release, this figure represents a 7.08 percentage point improvement over a base model trained with Alibaba’s Qwen3-14B.
"I gave the problem description to Cloud Code, it generated everything we built in an hour last year," Jana Dogan, a principal engineer at Google responsible for the Gemini API, wrote in a viral post on X last week that reflected the prevailing mood around the AI coding tool. Dogan was describing a distributed agent orchestration system that her team had spent a year developing — a system cloud code inferred from a three-paragraph prompt.
The comparison is instructive: While Anthropic’s Cloud Code has captured imaginations with demonstrations of end-to-end software development, Nous Research is betting that open-source alternatives trained on verifiable problems can close the gap — and that transparency in how these models are built matters as much as raw capability.
How Nous Research created an AI coding model that anyone can replicate
What sets the NousCoder-14B release apart from many competing announcements is its fundamental openness. Noos Research published not only the model weights but also the entire reinforcement learning environment, benchmark suite, and training harness built on the company’s Atropos framework – enabling any researcher with enough compute to reproduce or extend the work.
"Open-sourcing the Atropos stack provides the infrastructure needed for reproducible Olympiad-level logic research," An observer on X summarized the importance to the academic and open-source communities.
The model was trained by Joe Lee, a researcher at Nous Research and former competitive programmer. Lee’s technical report reveals an unexpectedly personal dimension: He compared the model’s improvement trajectory to his journey on the competitive programming platform Codeforces, where participants earn ratings based on competition performance.
Based on a rough estimate mapping LiveCodeBench scores to Codeforces ratings, Lee calculated that the NousCoder-14B’s improvement – from the approximately 1600–1750 rating range to 2100–2200 – represented a jump that required approximately two years of consistent practice between the ages of 14 and 16. The model was completed in four days.
"It was quite a surreal experience to watch that last training run," Lee wrote in the technical report.
But Li immediately noted an important caveat that speaks to broader questions about AI efficiency: They solved about 1,000 problems during those two years, while the model required 24,000. Humans, at least so far, remain dramatically more sample-efficient learners.
Inside the reinforcement learning system that trains on 24,000 competitive programming problems
The training process of NousCoder-14B provides a window into the increasingly sophisticated techniques used by researchers to improve AI reasoning capabilities through reinforcement learning.
The approach depends on what researchers say "Verifiable Rewards" – A system where the model generates code solutions, those solutions are executed against test cases, and the model receives a simple binary signal: true or false. This feedback loop, while simple conceptually, requires significant infrastructure to execute at scale.
Noos Research used a cloud computing platform model to run sandbox code execution in parallel. Each of the 24,000 training problems contains on average hundreds of test cases, and the system must verify that the generated code produces the correct output within time and memory constraints – 15 seconds and 4 gigabytes, respectively.
The training used a technique called DAPO (Dynamic Sampling Policy Optimization), which the researchers found performed slightly better than alternatives in their experiments. A major innovation involves "dynamic sampling" – Discarding training examples where the model either solves all attempts or fails all attempts, as these do not provide any useful gradient signal for learning.
Researchers also adopted "iterative reference expansion," Train the model with a 32,000-token context window before expanding to 40,000 tokens. During evaluation, expanding the context to approximately 80,000 tokens yielded the best results, reaching accuracy of 67.87 percent.
Perhaps most importantly, the training pipeline overlaps inference and validation – as soon as the model generates a solution, it begins work on the next problem while the previous solution is being examined. This pipelining, combined with asynchronous training where multiple model instances work in parallel, maximizes hardware utilization on expensive GPU clusters.
Growing data shortage could slow down progress of AI coding models
Li’s technical report hides a discovery with important implications for the future of AI development: the training dataset for NousCoder-14B includes "A significant portion of all readily available, verifiable competitive programming problems in a standardized dataset format."
In other words, for this particular domain, researchers are approaching the limit of high quality training data.
"The total number of competitive programming problems on the Internet is of approximately the same order of magnitude," Lee wrote, referring to the 24,000 problems used for training. "This suggests that within the competitive programming domain, we have reached the limit of high-quality data."
This observation reflects growing concern in the AI industry about data bottlenecks. While calculations continue to scale according to well-understood economic and engineering principles, training data is "increasingly finite," As Lee said.
"It appears that some of the most important research to be conducted in the future will be in the areas of synthetic data generation and data efficient algorithms and architectures," He concluded.
The challenge is particularly severe for competitive programming because the domain requires problems with known correct solutions that can be automatically verified. Unlike natural language tasks where human evaluation or proxy metrics suffice, the code either works or it doesn’t – making synthetic data generation significantly more difficult.
Lee identified a possible path forward: training models not only to solve problems but also to generate solvable problems, enabling a form of self-play similar to the techniques that proved successful in game-playing AI systems. "Once the synthetic problem arises, self-play becomes a very interesting direction," He has written.
The $65 Million Bet That Open-Source AI Can Compete With Big Tech
Nous Research has carved out a unique niche in the AI landscape: a company committed to open-source releases that compete with – and sometimes even surpass – proprietary alternatives.
The company raised $50 million in April 2025 led by Paradigm, the cryptocurrency-focused venture firm founded by Coinbase co-founder Fred Ehrsam. According to some reports, total funding reached $65 million. This investment reflects the growing interest in decentralized approaches to AI training, an area where Nous Research has developed its Psyke platform.
Previous releases include the Hermes 4, a family of models that we reported on "Outperforms ChatGPT without content restrictions," and DeepHermes-3, which the company said was the first "toggle-on logic model" – Allows users to activate extended thinking capabilities on demand.
The company has developed a distinctive aesthetic and community, leading to some skepticism as to whether style can trump substance. "Ofc I’m supposed to trust an anime php company. stop benchmarkmaxing ffs," A critic on X wrote, referring to Nous Research’s anime-style branding and industry practice of optimization for benchmark performance.
Others raised technical questions. "Based on benchmarks, Nemotron is better," one commenter said, referring to Nvidia’s language model family. Another asked if NousCoder-14B "Agent-centric or just ‘one-shot’ coding" – a distinction that matters for practical software development, where iteration on feedback usually gives better results than single attempts.
That’s what’s next, say researchers, as AI coding tools continue to improve
The release includes several directions for future work that indicate where AI coding research is headed.
Multi-turn reinforcement learning tops the list. Currently, the model only receives a final binary reward – pass or fail – after generating a solution. But competitive programming problems typically involve public test cases that provide intermediate feedback: compilation errors, incorrect output, deadline violations. Incorporating this feedback into multiple attempts can significantly improve training model performance.
Controlling the response length also remains a challenge. The researchers found that incorrect solutions tended to be longer than correct solutions, and that response lengths during training rapidly saturated the available context window – a pattern that various algorithm modifications failed to resolve.
Perhaps most ambitiously, Lee proposed "Problem formulation and self-play" – Training models for both solving and creating programming problems. This will directly address the problem of data scarcity by enabling models to generate their own training courses.
"Humans are great at generating interesting and useful problems for other competing programmers, but it appears that a significant gap still exists in LLM abilities in creative problem formulation," Lee wrote.
This model is now available on Hugging Face under the Apache 2.0 license. For researchers and developers who want to take the work further, Nous Research has published the complete Atropos training stack as well.
It required two years of teenage dedication for Lee to go from 1600-level newbie to 2100-rated competitor on Codeforces – an AI replicated in 96 hours. He needed 1,000 problems. The model required 24,000. But soon, these systems may learn to write their own problems, teach themselves, and leave human parameters behind entirely.
The question is no longer whether machines can learn to code. The question is whether they will soon be better teachers than us.
<a href