A small research lab backed by cryptocurrency venture capital trained a competitive AI model in just four days. It matches or beats much larger, more expensive systems from some of the biggest names in tech. And they published everything — the model weights, the training code, the benchmarks — for anyone to use, modify, or improve upon.
This is the story of NousCoder-14B, a programming model released by Nous Research that achieved a 67.87 percent accuracy rate on LiveCodeBench v6, a standard benchmark for competitive programming tasks. That represents a significant leap over its base model and puts it in the same conversation as proprietary systems costing many times more to develop.
Why Speed of Training Matters
The most striking detail is the training time. Nous Research trained the model on 48 of Nvidia’s B200 graphics processors over four days. To put that in perspective, many comparable models take weeks or months of training on larger clusters. The fact that this could be done in 96 hours suggests that AI development is getting faster and cheaper at a pace that’s hard to keep up with.
The model’s lead researcher, Joe Li, compared its improvement to his own journey as a competitive programmer. The jump in performance — from roughly a 1600 rating to a 2100-2200 rating on the Codeforces scale — took him nearly two years of practice. The model did it in four days. Li was quick to note that the model needed 24,000 training problems compared to the 1,000 he solved over those two years. Humans are still far more efficient learners. But the gap is narrowing fast.
What This Means for Irish Businesses
You might be wondering what a competitive programming model has to do with running a small business in Ireland. The answer is not about the model itself, but about what it represents. The cost of creating capable AI is falling rapidly. If a small research lab can train a competitive model in four days, the barriers to creating specialised AI tools are coming down.
For an Irish business, this suggests a future where you won’t need to pay subscription fees to a handful of big AI companies. Instead, you’ll be able to run specialised, open-source models that are fine-tuned for your specific needs — whether that’s analysing property data, optimising delivery routes, or generating quotes for construction jobs. The open-source model ecosystem is growing fast, and it’s becoming a genuine alternative to proprietary systems.
Think about what this means in practical terms. A small Irish engineering firm could take an open-source model, fine-tune it on Irish building regulations and pricing data, and have a specialist assistant that knows Irish construction law better than any general-purpose AI. No monthly subscription per user. No data sent to a US server. The model runs on your own hardware. That future is closer than most people realise.
The Challenge: Not Enough Data
Li’s technical report included a revealing observation: the training dataset for NousCoder-14B “encompasses a significant portion of all readily available, verifiable competitive programming problems.” In plain English, they’re running out of good training data for this particular domain. The researcher noted that “within the competitive programming domain, we have approached the limits of high-quality data.”
This is a challenge facing the entire AI industry. While computing power continues to scale, training data is increasingly finite. The solution, according to Li, lies in synthetic data generation and more data-efficient algorithms. For businesses, this means the race isn’t just about who has the biggest computers — it’s about who has the best data and knows how to use it.
What to Watch
For Irish businesses, the open-source AI trend is one to watch closely. The ability to run capable AI models without paying per-query fees to a US tech company is becoming real. If you’re already using AI tools, it’s worth keeping an eye on open-source alternatives — they’re catching up fast, and they offer something proprietary models don’t: control over your own data and costs.