NVIDIA's Nemotron Beats the Best Human at the Coding Olympics

A specialized version of NVIDIA's Nemotron model just outscored the highest-scoring human on a full International Olympiad in Informatics problem set, running under the exact same rules as the contestants in the room. The technical report puts the score at 535.4 out of 600, comfortably above the 361.12 gold threshold and above the 498.27 posted by the top human contestant.

The run happened in Uzbekistan alongside the official competition, with the International Technical Committee supervising. According to NVIDIA, the model had no internet access and faced the same time limits and submission constraints as the human contestants, using the same judging platform in parallel.

Why IOI is a brutal benchmark

IOI hands contestants a small number of algorithmic problems inside a strict time window, graded by a machine judge against hidden test cases. Solutions must hit tight runtime and memory limits, not merely produce correct output. That combination of precise specs and binary pass-or-fail feedback is close to ideal for training and evaluating a reasoning system, which is why competitive programming has become a key test of large language model reasoning. Previous standout results from DeepMind's AlphaCode line, OpenAI's reasoning models, and DeepSeek's Speciale variants reached or approached gold-medal thresholds. Surpassing the best human score is a different bar entirely.

Two models, one pipeline

The paper, Post-Training Language Models for Gold-Medal Performance, describes an end-to-end specialization pipeline built on top of NVIDIA's existing Nemotron base models. It produces two specialized variants:

• Nemotron-3-Nano-CC is a 30B-A3B model trained with both SFT and RL, meaning 30 billion total parameters with 3 billion active per token in a mixture-of-experts setup.
• Nemotron-3-Ultra-CC is a 550B-A55B model trained with SFT alone, skipping reinforcement learning entirely.

That split is deliberate. It isolates what supervised fine-tuning on high-quality reasoning traces can do at scale versus what a smaller model can extract from additional RL.

How the pipeline works

The authors curate 22,000 problems and use DeepSeek-V4-Flash to generate 1.2 million reasoning traces for the compact model and 477,642 for the larger one, excluding all evaluation problems from training. Starting from Nemotron-3-Nano-30B-A3B, they apply SFT and RL to produce Nemotron-3-Nano-CC.

The step doing the most work at inference time is a technique the authors call GenCorrect: a feedback-driven loop that iteratively generates, evaluates, and refines candidate solutions. In practice:

1. Generate a batch of candidate solutions in parallel.
2. Run them against available test signals, capturing failures.
3. Feed the failures back into the model and refine.
4. Repeat until the budget runs out or a solution passes.

This mirrors what a strong human contestant does with a debugger, automated across many candidates simultaneously. Inference becomes a generate-check-repair loop rather than a single forward pass.

What the numbers show

On IOI 2025, used as a retrospective benchmark, the gains from each stage are readable directly from the scores. Nano-CC improves from 130 points to 291 after post-training, then jumps to 468 with GenCorrect, clearing the gold threshold of 438.3. Ultra-CC reaches 502.

Setup	IOI 2025 Score
Nano base	130
Nano-CC (post-training)	291
Nano-CC + GenCorrect	468
Ultra-CC	502
IOI 2025 gold threshold	438.3

Two details stand out. GenCorrect alone moves the small model from 291 to 468, a roughly 60% jump from inference-time work with no additional training. The 30B-A3B model with the loop also lands within 34 points of the 550B model without it, which says something meaningful about where the compute should go.

The prospective IOI 2026 evaluation rules out the usual contamination complaint. The 535.4 score came from genuinely new, unseen problems under contest conditions, produced by a competition-specific Ultra-CC system guided by the retrospective results.

The reusable idea behind the score

A mid-sized MoE model plus a well-designed generate-evaluate-refine loop closed most of the gap to a 550B model. The leverage sits in the verifier and the loop, not raw parameter count.

IOI works so cleanly here because it ships with a perfect judge: hidden test cases, deterministic scoring, unambiguous specs. Real engineering tasks rarely offer that. When feedback is a flaky test suite, an ambiguous ticket, or a reviewer's preference, iterative refinement has much less to refine against.

Where this transfers

The pipeline maps well onto any task with a machine-checkable success signal:

• Code generation against strong unit test coverage.
• SQL generation with known schemas and expected result sets.
• Config or infrastructure code validated against a schema or type checker.
• Formal reasoning with automated proof checkers.
• Any domain where you can build a synthetic problem bank and generate reasoning traces from a stronger teacher model.

Open-ended software engineering, product decisions, anything requiring stakeholder context, and anything with noisy or delayed feedback sit outside this approach. The paper makes no claim to general software engineering ability, and the team built a competition-specific system for IOI 2026, not a general agent that happened to be good at the contest.

What's available

The work builds on NVIDIA's Nemotron research line. The base models it fine-tunes from, Nemotron-3-Nano and the earlier Nemotron-Cascade family, have been released with open weights and open post-training recipes on NVIDIA's Hugging Face page. The CC-specialized variants and GenCorrect implementation are described in the technical report; if the team follows the pattern set by Nemotron-Cascade 2, weight and dataset releases should accompany or follow the paper. Competitive-programming benchmarks are no longer a reliable ceiling for AI systems. The next honest test is what happens when the judge gets messier.