Nemotron 3 hits gold at IOI 2026 and IMO 2026

The two projects build on earlier work from the same teams: a previous IOI 2025 Hugging Face post showed test-time compute could push open-weight models to gold-level performance. The new results add a missing piece, namely that better specialization gives the inference system better candidates, critics, and refinements. For IOI, the progression on IOI 2025 makes the value of specialization concrete: Nano improved from 130 points before post-training to 280 after SFT and 291 after RL, then reached 468 with GenCorrect. For the stronger Ultra model, one SFT epoch was enough to outperform the fully post-trained Nano across IOI, ICPC, and LiveCodeBench Pro.

For IMO, the project trained one specialist with SFT and another with RL. The SFT checkpoint was strongest in the first search round, while the RL checkpoint achieved the best overall single-checkpoint result, and because their strengths were complementary the final system used both alongside the general model. A separate high-compute stage selected the final submission, and the entire system worked in natural language with no formal prover, external tools, or internet access.

The broader claim is that the medals came from co-designing the model, the data, and the inference loop, not from fine-tuning alone or brute-force sampling alone. Whether that recipe holds up as a general pattern will likely depend on whether other teams can reproduce the released checkpoints, datasets, and NeMo-Skills pipelines on tasks beyond these two competitions.