NVIDIA says separately specialized systems from its Nemotron 3 model family reached gold-medal score thresholds on the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO). The results are not one unchanged checkpoint winning two competitions: the coding system paired a specialist model with feedback-driven search, while the mathematics system combined three checkpoints with proof generation, verification and selection. 1 2
The company’s October 7 announcement brought together work already documented in September. The coding paper was first submitted on September 2 and the mathematics paper on September 9, so the announcement should not be read as two newly achieved October competition results. 1 3 4
Two scores, different participation status
For competitive programming, NVIDIA reports 535.4 out of 600, above its cited gold cutoff of 361.12 and highest human score of 498.27. But the company explicitly describes the run as an unofficial, unsupervised benchmark excluded from official IOI rankings. It was not an official IOI medal win. 5 1
For mathematics, NVIDIA’s paper reports official participation and grading that awarded 30 out of 42: full credit on Problems 1, 2, 4 and 5, and one point each on Problems 3 and 6. The IMO organizer’s results independently establish the human gold cutoff as 29; that source does not independently confirm the model’s score. 6 7
The coding paper describes two five-hour sessions, no internet, permitted local execution and a maximum of 50 submissions per problem, limited to one per minute. Researcher Igor Gitman reports human-matched time limits and no internet or external tools for the mathematics system. These restrictions describe time and tool constraints, not human-equivalent computing resources. 8 9
Fine-tuning was only part of the recipe
According to its model card, the coding specialist, Nemotron-3-Ultra-CC, has 550 billion total parameters and 55 billion active parameters. NVIDIA trained it for one supervised fine-tuning epoch on 477,642 synthetic reasoning traces, distilled from GLM-5.2 and covering 22,000 curated problems. At inference time, GenCorrect generates candidate programs, uses evaluation feedback and refines subsequent attempts within the submission budget. 5
The mathematics training took a different route. NVIDIA reports 414,890 filtered supervised examples across 15,818 proof problems, covering proof construction, refinement, verification and meta-verification. A separate reinforcement-learning specialist used 9,597 problems selected near the model’s capability boundary. 1
The released IMO recipe combines general-availability, supervised-fine-tuned and reinforcement-learning-specialized Nemotron 3 Ultra checkpoints. It uses eight generation prompts, a two-checkpoint verifier panel and up to seven refinement rounds, followed by a separate, higher-compute judging stage to select final proofs. This is a natural-language proof pipeline, not a formal theorem prover or a single-answer benchmark. 2
NVIDIA’s internal development results illustrate why the distinction matters. On its IOI 2025 development benchmark, the company reports Nano progressing from 130 points before post-training to 280 after supervised fine-tuning, 291 after reinforcement learning and 468 with GenCorrect, above that year’s reported 438.3 gold threshold. That progression supports the importance of inference-time search; the final gain cannot be credited to fine-tuning alone. It remains an internal comparison, not an independently controlled estimate of each component’s contribution. 1
Substantial compute—and imperfect verification
The reported results required substantial infrastructure. NVIDIA’s coding paper says deployment peaked at 760 GB300 GPUs. For mathematics, the company reports approximately 707 million generated tokens and 1,464 GB200 GPU-hours to obtain the submitted proofs. The complete mathematics competition run consumed approximately 2.31 billion tokens and 4,800 GPU-hours. Those figures do not establish an inexpensive route to reproducing the headline scores. 8 6
The mathematics paper also documents a verification failure. For the proofs available at the deadline, the internal verifier and a post-hoc model jury estimated roughly 32 points, compared with the reported official score of 30. The authors attribute the discrepancy to Problems 3 and 6 and identify a likely shared verification blind spot. Multiple model judges therefore did not eliminate correlated errors. 6
A later Problem 6 revision, produced after 8 hours 25 minutes of search, received 4/7 from independent human graders working without the official marking scheme. That implied an unofficial total of 33, but it does not replace the 30/42 competition result. 6
The headline coding score also came from one live run. In five post-competition experiments, NVIDIA’s authors report an average of 521.72, with scores ranging from 495.0 to 545.8. The range includes a result below the company’s cited highest human score. These repeats demonstrate variability within the authors’ testing, not independent external replication. 8
Released artifacts make the methods inspectable
NVIDIA has released the live coding checkpoint in NVFP4 format under the OpenMDW License Agreement 1.1. Its mathematics paper identifies released specialist checkpoints under OpenMDW-1.1 and training datasets and a benchmark under CC BY 4.0. The NeMo-Skills documentation links configurations, prompts and submitted IMO proofs, making important parts of the system available for inspection. 5 6 2
The work is not the first reported dual gold-level result within Nemotron. Earlier NVIDIA research on Nemotron-Cascade 2 already reported gold-level performance on both IMO and IOI 2025. Nor should raw scores across years become a direct capability ranking: the IMO human gold cutoff was 35 in 2025 and 29 in 2026, reflecting different competition problem sets. 10 7
Confirmation of the new results remains predominantly first-party. Deyron Labs published secondary coverage, but it attributes the results to NVIDIA and lists the company’s announcement as its sole source. That adds coverage, not independent experimental confirmation. The supplied research review, which closed on October 12, 2026, examined artifact pages and documentation but did not download weights or rerun the experiments. 11
Several questions remain unresolved: what organizer records can establish monitoring and human intervention during the unsupervised IOI run; what minimum hardware and expenditure can reproduce its score; whether the released recipes exactly match competition selection, quantization and runtime settings; and how often the mathematics verifiers share errors on fresh, independently human-graded problems. 1 8 2 6
The strongest implication is about system design: specialization, feedback, verification and selection worked together in NVIDIA’s reported competition results. The evidence does not, by itself, establish economical software engineering or reliable autonomous mathematical research. 2 6 8