Nemotron 3 Ultra reaches gold-medal score at IMO 2026

A paper posted to Hugging Face's Papers listing, credited to Ivan Moshkov and co-authors, lays out a full recipe for building a natural-language olympiad-mathematics proof system. Starting from the general-availability Nemotron 3 Ultra model, the authors trained two specialist checkpoints using supervised fine-tuning and reinforcement learning, and studied how checkpoint choice, verification, and refinement affect performance at test time.

The resulting test-time pipeline uses three Nemotron 3 Ultra checkpoints together, the general-availability model plus the two post-trained specialists, to run an iterative search: the checkpoints generate candidate proofs, verify them, and refine them, entirely in natural language, with no formal prover, no external tools, and no internet access. A separate high-compute stage then picks each final submission from the candidates the search produces; the source does not describe how that stage judges one proof over another.

Run against the 2026 International Mathematical Olympiad, the system scored 30 out of 42 points, reaching the gold-medal threshold. The paper does not state the numeric cutoff that defines that threshold, and it does not compare the score to other AI systems' or human competitors' results.

Alongside the paper, the authors released the two post-trained checkpoints, the supervised fine-tuning and reinforcement-learning data used to build them, the training and inference code for the search pipeline, the actual solutions the system submitted at IMO 2026, and a new benchmark, Nemotron-IMO-Bench, made up of 200 novel olympiad-level problems.

Key facts

The pipeline scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold.
Two specialist checkpoints were trained from the Nemotron 3 Ultra base model using supervised fine-tuning and reinforcement learning.
Three Nemotron 3 Ultra checkpoints, the general-availability model plus the two specialists, run an iterative generate-verify-refine search entirely in natural language, with no formal prover, external tools, or internet access.
A separate high-compute stage then selects each final submission from the candidates the search produces.
The authors released the two post-trained checkpoints, the training data and code, the submitted solutions, and a new 200-problem benchmark called Nemotron-IMO-Bench.

Why it matters

The paper spells out a complete training and inference recipe for olympiad-level mathematical reasoning, not just a final score. Starting from the general-availability Nemotron 3 Ultra model, the authors trained two specialist checkpoints with supervised fine-tuning and reinforcement learning, then built a natural-language, tool-free search pipeline around three checkpoints together. That pipeline reached 30 out of 42 points at IMO 2026, the gold-medal threshold, and the authors released the recipe itself: the checkpoints, the training data and code, the submitted solutions, and a new benchmark, so the result can be inspected and rebuilt rather than only cited.

Who it affects

Teams working on mathematical or verified reasoning, and specifically on test-time-compute search pipelines, get a full working reference rather than a single headline score: two post-trained checkpoints, the supervised fine-tuning and reinforcement-learning data used to build them, the inference code for the generate-verify-refine search, and the actual solutions the system submitted at IMO 2026. Anyone benchmarking models on hard mathematics also gets Nemotron-IMO-Bench, a new set of 200 olympiad-level problems the authors describe as novel.

How to use it

Everything needed to reproduce or build on the result has been released together: the two post-trained specialist checkpoints (the pipeline's third checkpoint is the already-public Nemotron 3 Ultra general-availability model), the supervised fine-tuning and reinforcement-learning data used to train them, the training and inference code for the iterative generate-verify-refine search and its separate high-compute selection stage, the solutions the system actually submitted at IMO 2026, and the new Nemotron-IMO-Bench set of 200 olympiad-level problems.

How solid is it

The 30-out-of-42, gold-threshold result is the authors' own reporting, not an independently verified score. What backs it up is that the underlying evidence is public: the two post-trained checkpoints, the training data and code, and the exact solutions submitted at IMO 2026 are all released, so the pipeline can in principle be rerun and checked rather than taken on trust. The paper has drawn light attention as of this writing: 6 points and a single comment on its Hugging Face listing.

Risks and caveats

Several details are left out of the source text. The exact numeric cutoff for the IMO gold-medal threshold is not given, so the margin behind the 30-out-of-42 score cannot be checked from the text alone. The separate high-compute stage that picks each final submission is described only as selecting the result, without any explanation of how it judges one candidate proof over another. The abstract says the team evaluated checkpoint choice, verification, and refinement, but the source gives no results for that comparison. There is also no comparison to other AI systems' or human competitors' IMO scores, and no co-author names or institutional affiliation beyond the model name Nemotron 3 Ultra.