NVIDIA reports Nemotron 3 Ultra-based systems reached gold-level scores at IMO 2026 (30/42) and an unofficial IOI 2026 run (535.4/600), and releases models, datasets and a benchmark

NVIDIA says systems built on its Nemotron 3 Ultra model scored 30/42 at IMO 2026 (gold threshold 29), marked by official IMO graders, using a natural-language generate-verify-refine loop with no tools or internet. It also reports 535.4/600 on the IOI 2026 problems (gold threshold 361.12, top human 498.27) in a live run under contestant constraints, but this was unofficial and is not in the official ranking. Alongside the results it released the Nemotron Labs IMO 2026 collection (SFT and RL checkpoints, two training datasets), Nemotron-IMO-Bench (200 olympiad problems), the Nemotron-3-Ultra-CC model on Hugging Face, and inference pipelines, prompts and submitted proofs in NeMo-Skills, plus IMO and IOI papers.

The main practical value is the open release. Researchers and developers can inspect, reproduce and build on a competition-grade reasoning recipe, and the new benchmark gives others a shared test. Enterprises should not read it as proof of general reliability, since olympiad problems are narrow, well-defined and heavily studied. Competitors face a higher bar for open models, but gold-level IMO performance by AI is no longer new, so the competitive shift is modest.

The IOI figure is self-run and unofficial, so comparing it to top human scores is suggestive rather than settled. The framing of Nemotron as a "strong, adaptable foundation" that turns into "world-class specialists" is marketing, and the results do not show the recipe transfers beyond contests. The IMO margin is thin (30 against a 29 threshold), and the coverage gives no cost, compute or number of attempts per problem.

1. 01Independent reproduction of the IMO and IOI results from the released checkpoints and NeMo-Skills pipelines, including compute cost and number of samples per problem.
2. 02Whether other labs adopt Nemotron-IMO-Bench and what scores open and closed models post on it over the next few months.
3. 03Evidence that the fine-tuning recipe works on non-competition domains such as enterprise coding, science or formal verification, and download and fork activity for the released checkpoints on Hugging Face.