BuildNot yet confirmed elsewhere1 publisher2 min readPublished
NVIDIA reused one Nemotron 3 base to reach gold-medal level at both IMO and IOI 2026
NVIDIA says fine-tuned Nemotron 3 models reached gold-medal level at IMO 2026 and IOI 2026, scoring 30 of 42 at the officially graded IMO. The company credits standard fine-tuning paired with a loop that generates and checks answers.
The Engineer · Build desk
What happened
- NVIDIA ran IOI 2026 live under the contestants' time, internet and submission rules, but the run was unofficial, unsupervised and excluded from the official ranking.
- For coding, NVIDIA curated 22,000 problems; the 30-billion-parameter Nano-CC got SFT and RL, and the 550-billion-parameter Ultra-CC got SFT.
- On IOI 2025, Nano climbed from 130 points to 291 through SFT and RL, then to 468 with the GenCorrect loop, past the 438.3 gold threshold.
- The IMO system worked entirely in natural language, with no formal prover, external tools or internet access.
Why it matters
- cost Copying the recipe commits a team to inference spend on every problem it solves, since the generate-and-check loop added more to Nano's IOI 2025 score than SFT and RL did.
- decision A team with a large base model has grounds to try a single SFT epoch before building an RL pipeline, at least for competitive coding.
- constraint Proof-style work with no test cases turns the checker into a training target of its own, as NVIDIA's verification and meta-verification data for IMO shows.
Nano's IOI 2025 run can be taken apart stage by stage. SFT added 150 points and RL added 11 [21][22]. The GenCorrect loop then added 177 [23]. The loop alone beat the 161 points both training stages produced together, and accounts for about 52% of Nano's 338-point climb [24]. Fully trained but without the loop, Nano sat 147.3 points short of gold [25]. NVIDIA writes that SFT produced most of Nano's gain [18], a statement that holds for the training stages alone. For those 177 points to show up on someone else's workload, that workload needs a way to score candidates that the refine step can trust.
Scale changed which stage mattered. One SFT epoch on the larger Ultra beat the fully post-trained Nano on IOI, ICPC and LiveCodeBench Pro, and NVIDIA says that finding shaped its IOI 2026 entry [1]. That system, Ultra-CC, scored 535.4 of 600 [2], about 89% of the available points [20]. Both coding models activate a tenth of their total parameters [26].
At IMO the models did their own checking. For each problem they generated candidate proofs, scored them, wrote critiques and refined the best attempts, and a separate high-compute stage picked the submission [15]. The IMO SFT corpus taught verification and meta-verification alongside proof writing: 414,890 examples over 15,818 problems [12], about 26 per problem [28]. The RL specialist used 9,597 problems chosen near the edge of what the model could solve [13]. In search, the SFT checkpoint led the first round and the RL checkpoint gave the best single-checkpoint result, so the final system ran both alongside the general model [14]. NVIDIA reports that mixing checkpoints was worth more than drawing more samples from one [19]. If each of the six problems carries 7 points, four full solutions account for 28 of the 30, leaving 2 points from the other two problems combined [27].
Both projects started from Nemotron 3 [3], and both gold systems ran on NVIDIA's own Ultra base [8][11]. "We did not need to build a new foundation model for every challenge," NVIDIA wrote [30]. The same post says "The training and inference runs were substantial, but the underlying approach is familiar and reproducible." [29] The post does not report GPU-hours or inference budgets, and every run in it was NVIDIA's own [3]. In my view the evidence supports one base model feeding several specialists for a team that already owns that base and the compute to post-train and search with it [3][29].
What to watch
- A supervised, official IOI evaluation of AI entrants that could confirm NVIDIA's self-run gold-level result.
- Whether NVIDIA publishes the CC and IMO specialist checkpoints or its proof training corpus for outside reproduction.
- A team outside NVIDIA reaching olympiad gold level by post-training an open base model on a disclosed compute budget.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Insufficient
- Hype gap+20
- Incentives75
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
For the stronger Ultra model, one SFT epoch was enough to outperform the fully post-trained Nano model across IOI, ICPC, and LiveCodeBench Pro. That finding guided the competition-specific Ultra-CC system used for IOI 2026.
ReportedSupportedSource: NVIDIA, Hugging Face blog2 sources— create a free account to open themView cited source - [2]
The Ultra-CC system used for IOI 2026 scored 535.4 out of 600.
ReportedSupportedSource: NVIDIA, Hugging Face blog2 sources— create a free account to open themView cited source - [3]
Starting from Nemotron 3, NVIDIA's teams used supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference to create systems that reached gold-medal level at both IMO 2026 and IOI 2026.
- [4]
The IOI result came from a live, prospective run under the same time, internet-access, and submission constraints as human contestants. It was an unofficial, unsupervised benchmark and was not included in the official IOI ranking.
- [5]
The IMO system's submitted proofs were graded by official IMO graders.
- [6]
For competitive programming, NVIDIA curated 22,000 problems and generated synthetic reasoning traces to train two specialists.
- [7]
Nemotron-3-Nano-CC, with 30 billion total parameters and 3 billion active parameters, received both SFT and RL.
- [8]
Nemotron-3-Ultra-CC, with 550 billion total parameters and 55 billion active parameters, received SFT.
- [9]
On IOI 2025, Nano improved from 130 points before post-training to 280 after SFT and 291 after RL.
- [10]
With GenCorrect, NVIDIA's iterative generate-evaluate-refine strategy, Nano reached 468 points on IOI 2025 and crossed the gold threshold of 438.3.
- [11]
For IMO, starting from Nemotron 3 Ultra, NVIDIA trained one specialist with SFT and another with RL.
- [12]
The IMO SFT corpus contained 414,890 quality-filtered examples across 15,818 unique proof problems, covering proof generation, refinement, verification, and meta-verification.
- [13]
The IMO RL model was trained on 9,597 proof problems selected near the model's capability frontier.
- [14]
The SFT checkpoint was strongest in the first search round, while the RL checkpoint achieved the best overall single-checkpoint result; the final system used both specialists alongside the general model.
- [15]
For each IMO problem, the models generated candidate proofs, scored them, produced critiques, and refined the most promising attempts. A separate high-compute stage selected the final submission.
- [16]
The entire IMO system worked in natural language, with no formal prover, external tools, or internet access.
- [17]
The IMO system scored 30 out of 42 points, including full credit on four of the six problems, and exceeded the official gold-medal threshold.
- [18]
SFT produced most of Nano's gain, with RL adding a smaller but consistent improvement.
- [19]
At IMO, using complementary SFT and RL checkpoints was more valuable than simply drawing more samples from one checkpoint.
- [20]
Ultra-CC's IOI 2026 score is about 89% of available points.
- [21]
SFT added 150 points to Nano's IOI 2025 score.
- [22]
RL added 11 points to Nano's IOI 2025 score.
- [23]
The GenCorrect loop added 177 points to Nano's IOI 2025 score.
- [24]
Training (SFT plus RL) added 161 points; the loop's 177 points are about 52% of Nano's total 338-point gain on IOI 2025.
- [25]
After SFT and RL but before GenCorrect, Nano was 147.3 points below the IOI 2025 gold threshold.
- [26]
Both coding specialists activate 10% of their total parameters.
- [27]
Assuming equal weighting of 7 points per problem, four full solutions are 28 of the 30 points, so the other two problems yielded 2 points combined.
- [28]
The IMO SFT corpus averaged about 26 examples per unique proof problem.
- [29]
The training and inference runs were substantial, but the underlying approach is familiar and reproducible.
- [30]
We did not need to build a new foundation model for every challenge.
Sources
1 independent publisher whose own reporting we read for this story.
- huggingface.coOne Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
1 article · October 7, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Test-time computeFollow
- Model Post-TrainingFollow
- AI mathematical reasoningFollow
- AI for competitive programmingFollow