NVIDIA Opens Nemotron Specialization Recipe: IOI and IMO Gold-Level Scores, Weights and Recipe Released
According to an NVIDIA official blog dated October 7, 2026, the team used Nemotron 3 as a base and, through SFT, RL and a generate-verify-refine inference loop, scored 535.4/600 at IOI 2026 and 30/42 at IMO 2026, with both reported as clearing the gold-medal line; the relevant checkpoints, datasets and evaluation pipelines are released on Hugging Face and NeMo-Skills.
What changed
According to an NVIDIA team official blog published on October 7, 2026, it built two specialized systems on a Nemotron 3 base. For competitive programming, Nemotron-3-Ultra-CC (550 billion total parameters, 55 billion active, after SFT) paired with a GenCorrect iterative generate-evaluate-refine strategy scored 535.4/600 in a live, unofficial IOI 2026 forward run not counted in official rankings, above the reported 361.12 gold line and the human top score of 498.27. The companion Nemotron-3-Nano-CC (30 billion total parameters, 3 billion active, after SFT and RL) was also used in control experiments; the report mentions Nano improving from 130 points before post-training to 280 after SFT, 291 after RL, and 468 with GenCorrect added. For math, a generate-verify-refine system was composed of Nemotron 3 Ultra's general, SFT and RL checkpoints; the SFT corpus contained 414,890 samples covering 15,818 distinct proof problems, RL training used 9,597 problems near the capability boundary, and the system operated entirely in natural language, without formal provers, external tools or the internet, ultimately scoring 30/42 as graded by official IMO graders. The blog also states that the relevant models, two training datasets, Nemotron-IMO-Bench (200 olympiad-level problems) and the inference pipelines and prompts in NeMo-Skills have been released.
What it means for developers
The direct implication for developers is that this result clarifies the division of labor among base model, post-training and inference loop: the report shows the same base can be adapted through different scales and post-training combinations to two difficult task types, code and proof, and the gains do not come from fine-tuning alone or sampling alone. If similar recipes are later copied into enterprise settings, the object of evaluation shifts from a single model output to the entire pipeline, where candidate generation, scoring, critique, refinement and final selection jointly determine the score. A distinction is needed between open weights and an open-source license: the blog says the models and data are released on Hugging Face and provides a reproducible pipeline, but the source does not state specific license terms, so being downloadable does not mean free for commercial use, and adopters must check the model license and data terms themselves. In addition, the IOI result is explicitly an unofficial, unsupervised benchmark run unrelated to official rankings, while the IMO result is stated to be graded by official graders, so the two are not equally comparable. On resources, the data show Nano with SFT and RL and Ultra with only one SFT epoch can surpass a fully post-trained Nano, suggesting the compute budget for specialization is not linear across scales, and fine-tuning costs and inference costs need to be accounted for separately.
Integration and evaluation
Suggestion one (conditional analysis, not a reported fact): for tasks requiring strict output formats or hidden-test pass rates, first verify whether a post-trained checkpoint plus feedback loop beats a pure prompting approach. The basis for choosing this is the source-reported large gain for Nano after post-training and further gain after adding GenCorrect, indicating the benefit comes from the combination of fine-tuning and multi-round refinement; but if your task cannot automatically construct a reliable judge, the premise of this route does not hold. Evaluation could be set as comparing, on the same set of self-built hidden tests, the base model, a single post-trained checkpoint, and three rounds of generate-evaluate-refine, in terms of pass-rate difference and per-task inference time multiple. Suggestion two: in math or long-chain reasoning scenarios, consider keeping both SFT and RL checkpoints and using complementary sampling, rather than increasing samples on a single checkpoint. The source says the SFT checkpoint is strongest in the first search round, the RL checkpoint is overall best as a single checkpoint, and the final system uses both; this requires your pipeline to support multi-model routing and unified scoring. Evaluation could be set as, under a fixed total compute budget, comparing repeated sampling from one checkpoint and complementary sampling from two checkpoints on a self-built proof set in terms of verification pass rate and the proportion falsely judged complete. Suggestion three: if reusing the NeMo-Skills pipeline, first confirm whether its dependencies and prompts match your model version, and prepare compute budget for the final selection stage, because the source mentions that stage alone consumes considerable compute. Evaluation could be set as measuring end-to-end latency and per-problem cost of the whole pipeline on the same hardware to confirm it can complete within business deadlines.
Limits and open questions
The source blog was published on October 7, 2026, and the competition results are self-reported by the NVIDIA team; the IOI part is explicitly an unofficial, unsupervised run not counted in official rankings, and although the IMO part is said to be graded by official graders, this article cannot independently verify the scores, problem versions or grading details. The source does not disclose the specific license terms for the Nemotron 3 checkpoints; open weights and an open license are not the same thing, and commercial use and redistribution require checking independently. Versions such as Nemotron-3-Ultra-CC currently have no exact match in the live KimiSeek catalog, so this version is not currently on KimiSeek, and the open access channels mentioned in the report refer to the vendor's own release locations; if third-party APIs or hosting services appear later, their prices and availability should be governed by the corresponding provider's pricing page, and this article cites no specific price figures. The two post-training and multi-round refinement routes described depend on domain-specific problems, high-quality reasoning traces and reliable judges, and the source does not say whether they are equally effective in general enterprise settings lacking such data; training and inference overhead is described as "not small" but no reusable cost data is given.
Illustrations from the original report
