Nous Research, the open-source artificial intelligence startup backed by crypto enterprise firm Paradigm, released a brand new competitive programming model on Monday that it says matches or exceeds several larger proprietary systems — trained in only 4 days using 48 of Nvidia's latest B200 graphics processors.
The model, called NousCoder-14B, is one other entry in a crowded field of AI coding assistants, but arrives at a very charged moment: Claude Code, the agentic programming tool from rival Anthropic, has dominated social media discussion since Recent Yr's Day, with developers posting breathless testimonials about its capabilities. The simultaneous developments underscore how quickly AI-assisted software development is evolving — and the way fiercely firms large and small are competing to capture what many imagine will turn out to be a foundational technology for the way software gets written.
type: embedded-entry-inline id: 74cSyrq6OUrp9SEQ5zOUSl
NousCoder-14B achieves a 67.87 percent accuracy rate on LiveCodeBench v6, a standardized evaluation that tests models on competitive programming problems published between August 2024 and May 2025. That figure represents a 7.08 percentage point improvement over the bottom model it was trained from, Alibaba's Qwen3-14B, in keeping with Nous Research's technical report published alongside the discharge.
"I gave Claude Code an outline of the issue, it generated what we built last yr in an hour," wrote Jaana Dogan, a principal engineer at Google liable for the Gemini API, in a viral post on X last week that captured the prevailing mood around AI coding tools. Dogan was describing a distributed agent orchestration system her team had spent a yr developing — a system Claude Code approximated from a three-paragraph prompt.
The juxtaposition is instructive: while Anthropic's Claude 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 built an AI coding model that anyone can replicate
What distinguishes the NousCoder-14B release from many competitor announcements is its radical openness. Nous Research published not only the model weights however the complete reinforcement learning environment, benchmark suite, and training harness — built on the corporate's Atropos framework — enabling any researcher with sufficient compute to reproduce or extend the work.
"Open-sourcing the Atropos stack provides the needed infrastructure for reproducible olympiad-level reasoning research," noted one observer on X, summarizing the importance for the tutorial and open-source communities.
The model was trained by Joe Li, a researcher in residence at Nous Research and a former competitive programmer himself. Li's technical report reveals an unexpectedly personal dimension: he compared the model's improvement trajectory to his own journey on Codeforces, the competitive programming platform where participants earn rankings based on contest performance.
Based on rough estimates mapping LiveCodeBench scores to Codeforces rankings, Li calculated that NousCoder-14B's improvemen t— from roughly the 1600-1750 rating range to 2100-2200 — mirrors a leap that took him nearly two years of sustained practice between ages 14 and 16. The model achieved the equivalent in 4 days.
"Watching that final training run unfold was quite a surreal experience," Li wrote within the technical report.
But Li was quick to notice a crucial caveat that speaks to broader questions on AI efficiency: he solved roughly 1,000 problems during those two years, while the model required 24,000. Humans, a minimum of for now, remain dramatically more sample-efficient learners.
Contained in the reinforcement learning system that trains on 24,000 competitive programming problems
NousCoder-14B's training process offers a window into the increasingly sophisticated techniques researchers use to enhance AI reasoning capabilities through reinforcement learning.
The approach relies on what researchers call "verifiable rewards" — a system where the model generates code solutions, those solutions are executed against test cases, and the model receives a straightforward binary signal: correct or incorrect. This feedback loop, while conceptually straightforward, requires significant infrastructure to execute at scale.
Nous Research used Modal, a cloud computing platform, to run sandboxed code execution in parallel. Each of the 24,000 training problems incorporates a whole bunch of test cases on average, and the system must confirm that generated code produces correct outputs inside time and memory constraints — 15 seconds and 4 gigabytes, respectively.
The training employed a method called DAPO (Dynamic Sampling Policy Optimization), which the researchers found performed barely higher than alternatives of their experiments. A key innovation involves "dynamic sampling" — discarding training examples where the model either solves all attempts or fails all attempts, since these provide no useful gradient signal for learning.
The researchers also adopted "iterative context extension," first training the model with a 32,000-token context window before expanding to 40,000 tokens. During evaluation, extending the context further to roughly 80,000 tokens produced the most effective results, with accuracy reaching 67.87 percent.
Perhaps most importantly, the training pipeline overlaps inference and verification — as soon because the model generates an answer, it begins work on the following problem while the previous solution is being checked. This pipelining, combined with asynchronous training where multiple model instances work in parallel, maximizes hardware utilization on expensive GPU clusters.
The looming data shortage that would slow AI coding model progress
Buried in Li's technical report is a finding with significant implications for the longer term of AI development: the training dataset for NousCoder-14B encompasses "a good portion of all available, verifiable competitive programming problems in a standardized dataset format."
In other words, for this particular domain, the researchers are approaching the boundaries of high-quality training data.
"The full variety of competitive programming problems on the Web is roughly the identical order of magnitude," Li wrote, referring to the 24,000 problems used for training. "This means that throughout the competitive programming domain, now we have approached the boundaries of high-quality data."
This statement echoes growing concern across the AI industry about data constraints. While compute continues to scale in keeping with well-understood economic and engineering principles, training data is "increasingly finite," as Li put it.
"It seems that a few of an important research that should be done in the longer term can be within the areas of synthetic data generation and data efficient algorithms and architectures," he concluded.
The challenge is especially acute for competitive programming since the domain requires problems with known correct solutions that will be verified routinely. Unlike natural language tasks where human evaluation or proxy metrics suffice, code either works or it doesn't — making synthetic data generation considerably tougher.
Li identified one potential avenue: training models not only to unravel problems but to generate solvable problems, enabling a type of self-play just like techniques that proved successful in game-playing AI systems. "Once synthetic problem generation is solved, self-play becomes a really interesting direction," he wrote.
A $65 million bet that open-source AI can compete with Big Tech
Nous Research has carved out a particular position within the AI landscape: an organization committed to open-source releases that compete with — and sometimes exceed — proprietary alternatives.
The corporate raised $50 million in April 2025 in a round led by Paradigm, the cryptocurrency-focused enterprise firm founded by Coinbase co-founder Fred Ehrsam. Total funding reached $65 million, in keeping with some reports. The investment reflected growing interest in decentralized approaches to AI training, an area where Nous Research has developed its Psyche platform.
Previous releases include Hermes 4, a family of models that we reported "outperform ChatGPT without content restrictions," and DeepHermes-3, which the corporate described as the primary "toggle-on reasoning model" — allowing users to activate prolonged pondering capabilities on demand.
The corporate has cultivated a particular aesthetic and community, prompting some skepticism about whether style might overshadow substance. "Ofc i'm gonna imagine an anime pfp company. stop benchmarkmaxxing ffs," wrote one critic on X, referring to Nous Research's anime-style branding and the industry practice of optimizing for benchmark performance.
Others raised technical questions. "Based on the benchmark, Nemotron is healthier," noted one commenter, referring to Nvidia's family of language models. One other asked whether NousCoder-14B is "agentic focused or simply 'one shot' coding" — a distinction that matters for practical software development, where iterating on feedback typically produces higher results than single attempts.
What researchers say must occur next for AI coding tools to maintain improving
The discharge includes several directions for future work that hint at where AI coding research could also be heading.
Multi-turn reinforcement learning tops the list. Currently, the model receives only a final binary reward — pass or fail — after generating an answer. But competitive programming problems typically include public test cases that provide intermediate feedback: compilation errors, incorrect outputs, closing date violations. Training models to include this feedback across multiple attempts could significantly improve performance.
Controlling response length also stays a challenge. The researchers found that incorrect solutions tended to be longer than correct ones, and response lengths quickly saturated available context windows during training — a pattern that various algorithmic modifications did not resolve.
Perhaps most ambitiously, Li proposed "problem generation and self-play" — training models to each solve and create programming problems. This may address the info scarcity problem directly by enabling models to generate their very own training curricula.
"Humans are great at generating interesting and useful problems for other competitive programmers, but it surely appears that there still exists a big gap in LLM capabilities in creative problem generation," Li wrote.
The model is available now on Hugging Face under an Apache 2.0 license. For researchers and developers who need to construct on the work, Nous Research has published the entire Atropos training stack alongside it.
What took Li two years of adolescent dedication to attain—climbing from a 1600-level novice to a 2100-rated competitor on Codeforces—an AI replicated in 96 hours. He needed 1,000 problems. The model needed 24,000. But soon enough, these systems may learn to jot down their very own problems, teach themselves, and leave human benchmarks behind entirely.
The query isn’t any longer whether machines can learn to code. It's whether or not they'll soon be higher teachers than we ever were.
