🔍 Read the full analysis: How To Fine-Tune Nemotron For IOI And IMO Gold-Level Results on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Hugging Face says specialized systems built from its Nemotron 3 models scored 535.4 out of 600 at IOI 2026 and 30 out of 42 at IMO 2026. IMO graders officially awarded the submitted proofs 30 points; the IOI score came from an unofficial run and was excluded from the competition ranking.
Hugging Face says two specialized systems built from its Nemotron 3 model family reached gold-threshold scores at the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO), as detailed in the original analysis. The IMO proofs received 30 of 42 points from official graders; the IOI system scored 535.4 out of 600 in an unofficial run that was not included in the competition’s official ranking.
For IOI, Hugging Face used a competition-specific version of Nemotron-3-Ultra-CC, trained with supervised fine-tuning (SFT) and paired with GenCorrect, a process that generates candidate solutions, evaluates them and iteratively revises them. The company reports that the system ran prospectively under the contest’s time, internet-access and submission constraints. Its 535.4 score was above the reported 361.12 gold threshold and the top human score of 498.27, but it was not an official entry or ranking result.
For IMO, the team combined the general Nemotron 3 Ultra model with SFT and reinforcement-learning checkpoints in a natural-language proof system. It generated possible proofs, scored and critiqued them, and revised selected attempts. Hugging Face says the submissions received full credit on four of six problems, for 30 points—above the stated gold threshold of 29. The system used no formal prover, external tools or internet access, according to the company.
The projects used distinct training material. Hugging Face reports that the IOI work drew on 22,000 programming problems and synthetic reasoning traces. The IMO SFT data contained 414,890 quality-filtered examples from 15,818 proof problems; the reinforcement-learning model was trained on 9,597 problems selected near the model’s capability frontier. These figures and scores are reported by the company.
Two Subjects, Two Kinds of Evidence
The reported results suggest that adapting a shared model for a specific discipline and giving it a structured way to check and revise answers can produce strong results on demanding contest tasks. The approach differs between subjects: IOI requires code that performs against hidden tests, while IMO requires written mathematical arguments that graders judge for correctness.
The evidence is not equally validated. IMO proofs were officially graded, while the IOI result was an unofficial benchmark run, despite the reported contest-like conditions. The scores are evidence about performance on these particular competitions; they do not by themselves show how the systems would fare across broader programming and mathematics work.
AI programming problem solving tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From IOI Experiments to Proof Writing
Hugging Face presents the 2026 work as an extension of experiments at IOI 2025, where test-time computation was used to improve open-weight model performance. The company reported that a Nemotron-3-Nano-CC score increased from 130 before post-training to 280 after SFT and 291 after reinforcement learning; GenCorrect then raised it to 468, above that year’s stated gold threshold of 438.3. An Ultra-CC version scored 502 with the same test-time strategy.
The 2026 projects applied related ideas to both coding and proof writing, but used separate specialist systems and training data. For IMO, Hugging Face says SFT and reinforcement-learning checkpoints had complementary strengths, prompting the team to combine them with the general model rather than rely on a single checkpoint. The results and descriptions in this article are based on the company’s account.
“Success at both points to something broader.”
— Hugging Face
machine learning model fine-tuning kits
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Checks Still Needed
The supplied account does not describe independent verification of the IOI score or explain how that run was audited. Although Hugging Face says it followed the competition’s time, internet and submission constraints, its unofficial status means it cannot be treated as an official IOI ranking or medal.
It is also not yet clear how well either system generalizes to other contests, unfamiliar proof styles or real-world programming tasks. The company’s reported results do not isolate how much of the performance came from fine-tuning, inference-time search, data selection or compute. Independent replication and external scrutiny would help establish how broadly the findings apply.
As an affiliate, we earn on qualifying purchases.
Checkpoints and Benchmarks for Review
Hugging Face says its Nemotron Labs IMO 2026 collection includes SFT and reinforcement-learning checkpoints, both training datasets and Nemotron-IMO-Bench, a set of 200 olympiad-level problems. The company also points to an IMO paper describing the training and generate-verify-refine system, as well as a NeMo-Skills repository.
Those materials could allow researchers to examine the methods and test the models on additional problems. The supplied source does not give a timetable for further releases or report an independent evaluation of the IOI system. Reproductions, especially of the unofficial IOI run, remain a key next step for judging how robust the reported performance is.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did a Nemotron system officially win an IOI medal?
No. Hugging Face reports a score of 535.4 out of 600, but says the run was unofficial and was not included in the IOI’s official ranking.
Was the IMO result officially graded?
Yes. Hugging Face says official IMO graders awarded the submitted proofs 30 of 42 points, above the stated gold threshold of 29.
How did the systems generate their answers?
The IOI system used supervised fine-tuning and GenCorrect to generate, evaluate and revise code solutions. The IMO system combined a general model with SFT and reinforcement-learning checkpoints, then generated, critiqued and revised natural-language proofs.
Have the reported results been independently reproduced?
The supplied source does not report independent replication. It also does not describe an external audit of the unofficial IOI run, so that result should be understood as a company-reported benchmark score.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
