NVIDIA's open Nemotron coder beats the top human at IOI 2026
It took IMO gold too, with no prover — and the 550B weights, the GenCorrect recipe, the datasets and a 200-problem benchmark all ship.
Copy markdown
535.4 at IOI, and the weights are on Hugging Face
Nemotron-3-Ultra-CC scored 535.4/600 on a live, unsupervised IOI 2026 run — past the 361.12 gold line and the 498.27 top-human score. The 550B-total / 55B-active MoE is downloadable now in NVFP4.
The real unlock is GenCorrect, and it ships
The jump comes from a test-time method, not raw scale: GenCorrect took the small Nemotron-Nano-CC from 130 to 468 points (gold is 438.3) on IOI 2025. The full pipeline and prompts are in the open NeMo-Skills repo, so you can port it to a model you can actually run.
Gold at the math olympiad too, no formal prover
A generate-verify-refine system on Nemotron 3 Ultra scored 30/42 at IMO 2026 — above the 29 gold threshold, with full marks on 4 of 6 problems — graded by official IMO graders, using no prover, external tools or internet.
A 200-problem benchmark to grade your own models
NVIDIA also released Nemotron-IMO-Bench (200 olympiad-level problems), both training datasets, and the SFT and RL checkpoints — enough to reproduce the result or stress-test your own reasoning stack.