The Pulse
NVIDIA Opens Nemotron’s IOI and IMO Gold Recipes
NVIDIA has released training materials and inference workflows behind Nemotron systems that reached gold-level scores at IOI and IMO 2026. The IMO system scored 30 of 42 points officially, while the IOI result was a prospective, unofficial
AI.info Team · Written with AI ·
“The system scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold.”
Ivan Moshkov et al., “An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics”
NVIDIA has published the training materials and inference methods behind Nemotron systems that reached gold-level performance in two very different contests: the International Olympiad in Informatics and the International Mathematical Olympiad. The release, published October 7 on Hugging Face, brings together research papers, model checkpoints, datasets, benchmarks and code for reproducing the approaches.
The results are not equivalent in status. Nemotron’s 30-point IMO score was an official competition result, graded by official IMO graders; the IOI score came from a prospective, unsupervised evaluation and was not part of the official ranking. NVIDIA presents the pair as evidence that a shared foundation model can be adapted for distinct specialist tasks through post-training and carefully designed inference.
Two contests, two different scorecards
At IOI 2026, Nemotron-3-Ultra-CC scored 535.4 out of 600 under the same time, internet-access and submission constraints as human contestants, NVIDIA says. The score exceeded the stated gold threshold of 361.12 and the top human score of 498.27. The company explicitly describes the run as an unofficial benchmark, not an official IOI placement.
At IMO 2026, the system earned 30 of 42 points, one point above the official gold cutoff of 29. It received full credit on Problems 1, 2, 4 and 5, and one point each on Problems 3 and 6. The submission system used three Nemotron 3 Ultra checkpoints and generated, checked and revised natural-language proofs without a formal prover, external tools or internet access.
Fine-tuning paired with answer search
The IOI work starts from 22,000 curated programming problems. NVIDIA trained Nemotron-3-Nano-CC, a 30-billion-parameter model with 3 billion active parameters, using supervised fine-tuning and reinforcement learning; Nemotron-3-Ultra-CC, with 550 billion parameters and 55 billion active, received supervised fine-tuning.
At inference time, the IOI system uses GenCorrect, which generates candidate solutions, evaluates them and revises them using feedback. The paper’s earlier IOI 2025 experiments show how the pieces interact: Nano-CC rose from 130 points before post-training to 291 after it, then reached 468 with GenCorrect, above that year’s 438.3 gold threshold. Ultra-CC scored 502 with the same inference strategy.
For IMO, NVIDIA trained one specialist checkpoint with supervised fine-tuning and another with reinforcement learning. The final pipeline combined both with the general Nemotron 3 Ultra model: the checkpoints proposed proofs, assessed candidates and supplied critiques for revision, while a separate, higher-compute stage selected the proofs submitted to the contest.
What NVIDIA has released
The Hugging Face announcement links to IMO specialist checkpoints, training datasets and Nemotron-IMO-Bench, a collection of 200 olympiad-level problems. NVIDIA says its NeMo-Skills repository contains the IMO inference pipeline, prompts, submitted proofs and a reproducible quickstart. For coding, the Nemotron-3-Ultra-CC model, IOI paper, evaluation and inference pipeline are also available.
The published IMO paper adds a measure of the resources behind the result. All six submitted proofs were found within about 707 million generated tokens and 1,464 GB200 GPU-hours; finishing the search rounds already in flight brought the full competition run to about 2.31 billion tokens and 4,800 GPU-hours. Letting the Problem 6 search run past the contest cutoff, outside the official result, took another 610 million tokens and 890 GPU-hours. NVIDIA’s releases therefore expose more than model weights, but reproducing the competition result still entails a substantial search and compute budget.
Gold scores do not erase the verification gap
The IMO paper identifies a concrete weakness in model-based checking. At the contest cutoff, the system’s search-time verifier and an independent model jury, run after the contest, both put the score at roughly 32 points, two more than the official 30. Both credited the submitted proofs for Problems 3 and 6, which each received only one official point. The authors say the error appears to reflect a shared blind spot in model-based verification.
That gap matters because the released recipe relies on models to judge candidate work before submission. The contest outcomes show that Nemotron can produce strong results in coding and proof-writing, while the IMO discrepancy shows why an internal verifier’s confidence cannot substitute for independent grading.