Nemotron hit IMO 2026's gold threshold at 30/42, without a formal prover

The abstract of "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics" states that the system scored 30 out of 42 points at IMO 2026 and reached the gold-medal threshold. That is all the paper's Hugging Face page says about the number. It does not name the graders, does not say whether the score came from the competition's own marking or from an internal process, and does not break the 30 points down by problem. Treat it as the authors' report until someone outside the group reproduces the run.

The page is more specific about the conditions. The pipeline works entirely in natural language: no formal prover, no external tools, no internet access.

Three checkpoints, two training stages, one selection pass

Everything starts from Nemotron 3 Ultra. The team took the general-availability model and post-trained two specialist checkpoints on top of it, drawing on supervised fine-tuning and reinforcement learning. That puts three models in the search loop instead of one, the same appetite for small, specialized open models that shows up in Nvidia's Nemotron 3.5 Lightning.

At inference those checkpoints drive an iterative cycle: generate a candidate proof, verify it, refine it, repeat. A separate stage with a much heavier compute budget then picks the final submission. Splitting selection from generation is a budgeting decision. Most of the time goes to cheap iteration, and the expensive pass is saved for the last call. Routing policies are converging on the same idea: CoBa matches best-of-16 majority voting with 58.9% fewer tokens by balancing compute across candidates.

The abstract calls the result an open-model test-time-compute pipeline, which puts as much weight on inference design as on training. It does not say how many candidates the search generates per problem, how the verifier scores them, or what the selection stage optimizes for. Those details decide whether the recipe transfers. Test-time scaling is not free money; sequential reasoning has been hitting diminishing or even negative returns, which is why work like ThinkRetrieve argues for injecting retrieval mid-reasoning rather than tacking on more steps.

What the release contains

The paper ships more than a write-up.

Artifact	What it covers
Two post-trained checkpoints	The specialist models the search runs on
Training data	Published as nvidia/Nemotron-Math-Proofs-v3-SFT and nvidia/Nemotron-Math-Proofs-v3-RL
Training and inference code	The pipeline itself
Submitted solutions	The system's answers to the IMO 2026 problems
Nemotron-IMO-Bench	200 novel olympiad-level problems

Two items carry more weight than the rest. The submitted solutions are the system's real answers to the competition problems, which is what makes an outside audit possible at all. Nemotron-IMO-Bench, meanwhile, adds 200 problems the authors describe as novel, giving the next team a test set that has not already been absorbed into training data.

On Hugging Face, the checkpoint nvidia/Nemotron-3-Labs-Ultra-Math-RL appears as a model citing the paper, and the two training sets appear as nvidia/Nemotron-Math-Proofs-v3-SFT and nvidia/Nemotron-Math-Proofs-v3-RL. That is a genuine release. It is not yet a reproduced one. Between publishing a recipe and confirming it works outside the authors' own setup there is a gap nobody has closed. Open-weight releases keep arriving faster than independent reproductions, a pattern visible across the sub-200M model boom as well.

Why natural language is the risky half

Formal proof assistants turn an argument into something a machine can check line by line. This pipeline refuses that route. It writes and reasons in ordinary mathematical prose, with no prover to fall back on and nothing external to consult.

The trade is legibility for certainty. Natural-language proofs read the way a human solution reads, and the framing suggests that was the point. But the verification step sits inside the pipeline being evaluated, so the system is partly grading its own work. A machine-checkable certificate would not have that problem. A language-model verifier does. Tools that look rigorous in isolation can still reach the wrong conclusion, as research on agents that run flawless statistics found.

The abstract does not describe what the verification stage checks, how often it rejects a correct proof, or how many refinement rounds a typical problem needs. Without those numbers, the gold-medal claim rests on a scoring process the reader cannot inspect.

What an open gold-level pipeline changes

If the result holds, the significance sits in what got released rather than in the score. Weights, training data, code and solutions together let another lab attempt the same run without rebuilding it from scratch.

That lands in a benchmarking field already drifting away from static tests. Fixed question sets have lost much of their diagnostic power as models train through them, and the field has moved toward dynamic evaluations and expert-level exams. Against that backdrop a self-reported olympiad figure is harder to place than it once was. It is one strong number from one team, on one paper, with the evaluation stack built by the same people who built the system.

The release is the part worth watching. Whether 30 out of 42 survives an outside audit is still open.