How NVIDIA trained Nemotron 3 to hit Olympiad gold

The company released open weights and its post-training recipe after beating the top human score in the IOI and reaching gold thresholds in the IMO.

NVIDIA published its complete post-training methodology and released model weights for the Nemotron 3 family, demonstrating performance that hit gold-medal standards in two demanding academic competitions: the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO). Technical details appeared on the Hugging Face blog and in arXiv papers, alongside open code in the NeMo-Skills repository.

High-stakes benchmarks in code and mathematics

The programming contest results stood out immediately. The specialized Nemotron-3-Ultra-CC model posted 535.4 out of 600 possible points on the IOI 2026 problem set. That score cleared the gold-medal cutoff of 361.12 points and topped the 498.27 points scored by the highest-performing human student in the contest. NVIDIA ran this evaluation prospectively under competition time limits, though the system operated in a parallel benchmark rather than entering the official student award ranks.

In the IMO 2026 evaluation, the system achieved 30 out of 42 points, clearing the 29-point gold cutoff. Unlike competing projects such as Google DeepMind’s AlphaProof, which translate problems into formal verification languages like Lean or Isabelle, Nemotron operated strictly in standard natural language. Without web search or external mathematical tools, it earned full marks (7 out of 7) on four of the six problems, while receiving 1 out of 7 on the remaining two, as evaluated by official competition markers.

The 4-stage Nemotron 3 recipe

To achieve these results without formal verification solvers, NVIDIA structured its post-training workflow into four sequential phases:

Base model: The foundation relies on Nemotron 3, a mixture-of-experts (MoE) architecture with 550 billion total parameters and 55 billion active per token.
Dense synthetic data: Step-by-step reasoning traces were filtered and generated, producing 22,000 curated programming tasks and roughly 414,000 mathematical solutions.
Fine-tuning and reinforcement: Standard supervised fine-tuning (SFT) paired with reinforcement learning via the NeMo-RL library.
Test-time computation: Rather than relying on a single output attempt, the system conducts search and automated corrections during generation.

The hardware footprint and GenCorrect

The decisive performance boost occurred in the fourth stage through structured test-time compute. For competitive programming, NVIDIA introduced an algorithm called GenCorrect. It generates up to 200 candidate programs in parallel, filters them for algorithmic diversity, and applies iterative debugging based on feedback from local test cases on subtasks.

That dynamic search requires substantial hardware. During the IOI evaluation, running these parallel inference passes consumed up to 760 NVIDIA GB300 GPUs. The gold-level result was not produced by raw weight quality alone; it depended heavily on massive parallel search capable of testing, failing, and revising code before settling on a final submission.

Open weights and industry pressure

Alongside its technical documentation, NVIDIA released weights for Nemotron-3-Ultra-CC and its math-focused variants on Hugging Face, along with the Nemotron-IMO-Bench dataset featuring 200 unseen problems. Code for the test suites was added to NeMo-Skills.

By releasing these artifacts publicly, NVIDIA puts pressure on competitors that keep their reasoning models behind proprietary APIs. The findings confirm that recent progress in machine reasoning relies just as much on computing power spent while answering a question as on the sheer volume of data ingested during pre-training.