Nemotron reaches Olympiad gold level: NVIDIA combines fine-tuning with solution verification

The gold medal cutoff was 361.12 points, while the top human competitor scored 498.27. NVIDIA ran the system during the competition under its constraints. GenCorrect adds solution checking and correction to code generation.

Limited fine-tuning. On IOI 2025 tasks, Nemotron-3-Nano-CC scored 130 points before training, 280 after training on existing solutions, and 291 after reinforcement learning. With GenCorrect, the score rose to 468 points against a gold medal cutoff of 438.3. In this experiment, the system crossed the gold medal cutoff after checking and correction were added.

NVIDIA trained Ultra-CC in a single pass over 477,642 synthetic solution walkthroughs from GLM-5.2. They covered 22,000 problems from 16 competition families. The IOI model is intended for competitive programming and research into code reasoning.

At IMO 2026, NVIDIA used the general-purpose Ultra and two math models fine-tuned using different methods. The system scored 30 out of 42 points against a gold medal cutoff of 29 and earned full marks on 4 of the 6 problems. Official IMO graders checked the proofs.

The math system also relies on repeated verification. In the first round, it generates 384 proof attempts, checks each one 16 times using math models, and gives each finalist 48 evaluations before selecting an answer. Generations reach 512,000 tokens.

Launch instructions are available. The Nemotron Labs IMO 2026 collection includes the general-purpose Ultra, Nemotron-3-Labs-Ultra-Math-SFT, and Nemotron-3-Labs-Ultra-Math-RL. NVIDIA also published two training datasets and Nemotron-IMO-Bench, containing 200 problems. Running the math pipeline in NeMo-Skills requires three models served through OpenAI-compatible endpoints.

The input file must use JSONL format with `problem` and `problem_idx` fields. Copy `configs/imo2026-ensemble.yaml` to `my-run.yaml`, specify the endpoints and models, then run `python recipes/nemotron-imo-tts/run.py --config my-run.yaml --output-dir runs/my-run`. The `--dry-run` flag checks the configuration, and completed proofs are saved in `submissions.jsonl`.

For an initial test, NVIDIA suggests `configs/smoke.yaml`, which uses fewer requests. The IOI checkpoint NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 was published on Sep 3, 2026. Its model card includes a Docker command for four GB300 GPUs using `vllm/vllm-openai:v0.22.0` and an example API call.

Original source: [NVIDIA on Nemotron's IOI and IMO results, Oct 7, 2026](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026).

The IOI model has 550 billion parameters, activates 55 billion per request, and supports a context window of 262,144 tokens.