NVIDIA details the Nemotron recipe behind gold-level IOI and IMO systems
A new NVIDIA case study connects two September releases, showing how supervised fine-tuning, reinforcement learning and iterative verification produced gold-level results in competitive coding and mathematics—with important limits on what the scores establish.
2 min read
NVIDIA has published a consolidated account of how it specialized Nemotron 3 for two very different olympiad tasks: competitive programming at IOI 2026 and natural-language proof writing at IMO 2026. The post does not introduce a new base model—the cited checkpoints and papers were released in September—but it makes the shared recipe clearer: curate difficult domain data, apply supervised fine-tuning and reinforcement learning where useful, then combine the specialist with an inference loop that generates, critiques and refines answers.
For IOI, NVIDIA reports that Nemotron-3-Ultra-CC scored 535.4 out of 600, above the 361.12 gold threshold and the top human score of 498.27. The run was prospective and used the same time, internet-access and submission constraints as contestants, but it was unofficial, unsupervised and excluded from the official ranking. The result also used GenCorrect, a multi-round system that evaluates candidate programs and feeds results into later attempts—not a single-pass model score.
For IMO, an ensemble of the general Nemotron 3 Ultra model plus SFT and RL math checkpoints generated, verified and refined proofs without internet access, external tools or a formal prover. Official graders awarded 30 of 42 points, one above the gold threshold. Here too, the system-level result should not be read as proof of broad model superiority.
The downloadable coding checkpoint has 550 billion total parameters, 55 billion active parameters, a 262,144-token context window and an OpenMDW 1.1 license; NVIDIA’s example deployment uses four GB300 GPUs. The BF16 math checkpoint is the same scale, supports up to one million tokens, and recommends eight B200 GPUs with roughly 1.5TB aggregate HBM. Neither checkpoint was hosted by a Hugging Face inference provider when checked.
Sources
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.