News · 2026-09-11

NVIDIA publishes the whole recipe behind an IMO gold score, weights and all

NVIDIA researchers have published the complete recipe behind a Nemotron-based system that scored 30 out of 42 at the 2026 International Mathematical Olympiad, one point above the gold-medal cutoff, as marked by the olympiad's official graders. The paper, An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics, posted on 9 September 2026, ships the model checkpoints, training data, a new 200-problem benchmark, the inference code and the proofs actually submitted.

Key facts

• 30 of 42 points from an official IMO 2026 entry, above the gold cutoff of 29.
• The system "operates entirely in natural language, with no formal prover, external tools, or internet access."
• Two specialist checkpoints are released, each a 1.12 TB download; NVIDIA's model card recommends at least eight B200 GPUs, about 1.5 TB of combined GPU memory.
• Primary source: the paper on arXiv, by Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi and Igor Gitman.

The opposite of a rushed announcement

The same week that 25 Fields Medallists complained about AI results "announced in a rush, leaving no time for a proper writeup" (our story), an NVIDIA team did the reverse: a detailed method, full compute accounting, and every artefact needed to check the claim. The International Mathematical Olympiad is the world's leading competition for secondary-school mathematicians. Contestants write full proofs to six problems over two days, each worth seven points, and gold is the top tier of medals.

How it works

The system starts from Nemotron 3 Ultra, a mixture-of-experts model with 550 billion parameters, of which 55 billion are active for any given word. The team trained two specialists from it: one by supervised fine-tuning on 414,890 examples whose proofs were generated by DeepSeek-V4-Pro, and one with reinforcement learning.

At contest time the approach is mostly about spending compute while answering. The checkpoints produced 384 proof attempts per problem. Each specialist judged every proof eight times, and a proof was accepted only if all sixteen judgments agreed. Rejected proofs were revised for up to eight rounds, and the finalists were ranked by 48 olympiad-style grades. Picture a student writing hundreds of drafts, showing each to two strict tutors eight times apiece, and handing in only the ones both tutors approved every single time. Finding the proofs that were submitted took about 1,464 GPU-hours on NVIDIA's GB200 chips; the whole run took about 4,800.

The paper is also honest about what happened after the bell. The system kept searching past the four-and-a-half-hour contest cutoff, and after more than eight hours it produced a new proof for the hardest problem, Problem 6. Its own verifier did not accept that proof, yet it scored 4 out of 7 in what the authors call "an unofficial independent human regrade," at a cost of 890 extra GPU-hours. The official result stays at 30. Keeping the two numbers separate is exactly the kind of bookkeeping the mathematicians' declaration asks for.

What "open" means here

The release includes both specialist checkpoints under NVIDIA's OpenMDW licence, the supervised and reinforcement-learning training data, the inference code on GitHub, the reinforcement-learning recipe, and a new benchmark of 200 olympiad problems written with the veteran problem-setter Titu Andreescu. Everything is gathered in a Hugging Face collection, and the model card lists the hardware.

That matters because a claim you can rerun is a different kind of claim. Mathematicians can read the submitted proofs line by line, and other labs can test whether the recipe transfers.

The honest caveats

The paper is unusually candid about its own blind spot. The team's model-based graders scored the run at about 32, two points above the official 30, which the authors call "a shared blind spot in model-based verification." They add: "accepted means that a proof satisfies this internal verification criterion; it does not by itself imply correctness under independent or official grading." This is the core weakness of AI grading AI: two judges trained on similar material can make the same mistake.

Open is also not the same as runnable at home. Each checkpoint is a 1.12 TB download, and the model card's minimum recommendation is a node of eight B200 GPUs with about 1.5 TB of combined memory, or at least eight H100, H200, GB200 or GB300 GPUs spread across several machines. And the training data leans on proofs written by another lab's model, DeepSeek-V4-Pro, so the recipe inherits whatever that model got wrong. Unlike systems that write proofs in a proof assistant such as Lean, nothing here is machine-checked; the guarantee is the olympiad graders' marking.

Primary source, verified: read the paper → (arXiv 2609.10712)

Ground Truth. (2026, September 11). NVIDIA publishes the whole recipe behind an IMO gold score, weights and all. Ground Truth. https://groundtruth.day/news/nvidia-publishes-the-whole-recipe-behind-an-imo-gold-score.html

@misc{groundtruth:nvidia-publishes-the-whole-recipe-behind-an-imo-gold-score,
  title  = {NVIDIA publishes the whole recipe behind an IMO gold score, weights and all},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/nvidia-publishes-the-whole-recipe-behind-an-imo-gold-score.html}
}

Topics: mathematics · nvidia · open-weights · reasoning · rl-post-training · research