Nvidia researchers report that a competition specific system built on their open Nemotron 3 Ultra model scored 535.4 of 600 at the 2026 International Olympiad in Informatics, above the gold threshold of 361.12 and above the top human contestant's 498.27, in a prospective run under the same time, internet access and submission constraints as the students.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source The system, called Nemotron-3-Ultra-CC, is a 550 billion parameter mixture of experts with 55 billion active parameters, fine tuned on 22,000 curated problems and driven at test time by a generate, evaluate and refine loop the authors call GenCorrect.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source 2 Ken Ashe 2026-09-03 Ultra-CC 550B total with 55B active; Nano-CC 30B with 3B active; GenCorrect 291 to 468 with no additional training; competition specific system; single prospective run; unspecified compute budget; machine checkable problems differ from real engineering. Open source It matters because prior systems had cleared gold thresholds without beating the best human on a full problem set, and because the base model is open.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source Our assessment, with high confidence, is that the score is genuine; with moderate confidence, that the method rather than the model is the transferable asset; and with low confidence that it says anything about general software engineering.

How the score was produced

The pipeline has two halves. The training half curated 22,000 competitive programming problems and generated synthetic reasoning traces, 1.2 million of them distilled from DeepSeek-V4-Flash for the smaller Nano model and 477,642 for Ultra, then applied supervised fine tuning, with reinforcement learning added only for Nano.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source 3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source The test time half, GenCorrect, generates many candidate solutions in parallel, runs them against test cases, feeds the errors back and refines, over repeated rounds.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source

The retrospective numbers on IOI 2025 show where the gains come from. Nano-CC went from 130 points as a base to 291 after post training, then to 468 with GenCorrect and no additional training, clearing that year's 438.3 gold line; Ultra-CC reached 502.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source The largest single jump, 291 to 468, came from spending more compute at inference, not from a better model.2 Ken Ashe 2026-09-03 Ultra-CC 550B total with 55B active; Nano-CC 30B with 3B active; GenCorrect 291 to 468 with no additional training; competition specific system; single prospective run; unspecified compute budget; machine checkable problems differ from real engineering. Open source The IOI 2026 run then used a competition specific configuration, executed live in Uzbekistan under supervision of the official International Technical Committee on the contest's own judging platform.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source

Sourcing caveat: every number here comes from the authors' preprint and secondary coverage of it. The result is a single prospective run that no third party has replicated, and the compute budget of the GenCorrect loop is not disclosed.2 Ken Ashe 2026-09-03 Ultra-CC 550B total with 55B active; Nano-CC 30B with 3B active; GenCorrect 291 to 468 with no additional training; competition specific system; single prospective run; unspecified compute budget; machine checkable problems differ from real engineering. Open source

Why the open base matters

Nemotron 3 Ultra was released on 4 June 2026 as an open model with a hybrid Mamba attention mixture of experts design, NVFP4 pretraining and a 1 million token context.4 NVIDIA Research 2026-06-04 Base model released 4 June 2026; 550B total, 55B active; hybrid Mamba attention mixture of experts; NVFP4 pretraining; 1 million token context; throughput claims of 5.9, 4.8 and 1.6 times against GLM-5.1, Kimi-K2.6 and Qwen-3.5. Open source That the competition system was built on an open base, rather than on a frontier lab's proprietary model, changes who can attempt this. A lab with the base weights, a problem corpus and enough inference compute can reproduce the recipe. It also serves Nvidia's commercial interest directly: GenCorrect converts GPU hours into points, and the paper is, among other things, a demonstration that inference compute is the lever.

Who gains and who loses

Nvidia gains twice: a headline result for its open model family and an argument that test time compute, which runs on its hardware, is where capability now comes from.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source DeepSeek gains a quiet credit, since its V4-Flash model generated the traces that taught Nano to reason.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source Teams building on open weights gain a documented recipe for turning a general model into a specialist.

The losers are narrower than the headline suggests. Competitive programming as a benchmark loses signal: once a specialized open system beats the top human live, the contest stops discriminating between frontier systems.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source Proprietary labs lose a talking point, since gold level performance no longer requires their models. Human contestants lose nothing that matters; the IOI is a contest for students, and the result changes what the score means for AI rather than for them.

The counter case

The strongest objection is the one the authors and the commentary both make: this is a competition specific system, not a general agent that happened to be good at the contest.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source IOI problems are closed, precisely specified and machine checkable, which makes them close to the ideal shape for a generate and verify loop and far from the shape of real engineering work, where the specification is the hard part and there is no oracle to grade the output.2 Ken Ashe 2026-09-03 Ultra-CC 550B total with 55B active; Nano-CC 30B with 3B active; GenCorrect 291 to 468 with no additional training; competition specific system; single prospective run; unspecified compute budget; machine checkable problems differ from real engineering. Open source The 535.4 therefore says a great deal about GenCorrect on verifiable tasks and very little about Devin style agents on a codebase. Two further weaknesses: a single run leaves the variance unknown, and an unspecified compute budget leaves the cost per point unknown. If the loop consumed compute worth far more than a contestant's laptop, the comparison to a human is a comparison of budgets, not intelligence.

What to watch

  • Independent replication. A third party reproducing a gold level score with the released recipe on IOI 2025 or 2026 problems within six months would confirm the method; failure to replicate would move the result into the single lab claim category.1 arXiv 2026-09-02 Five authors; submitted 2 September 2026; 22,000 problems; Nano-CC 30B with SFT and RL, Ultra-CC 550B with SFT only; GenCorrect test time loop; IOI 2025 path 130 to 291 to 468 for Nano-CC and 502 for Ultra-CC; IOI 2026 score 535.4 of 600 against gold 361.12 and top human 498.27; same constraints as human competitors. Open source
  • A disclosed compute budget. If Nvidia publishes GPU hours or dollar cost for the IOI 2026 run by the end of 2026, the cost per point becomes comparable across systems; continued silence keeps the efficiency question open.2 Ken Ashe 2026-09-03 Ultra-CC 550B total with 55B active; Nano-CC 30B with 3B active; GenCorrect 291 to 468 with no additional training; competition specific system; single prospective run; unspecified compute budget; machine checkable problems differ from real engineering. Open source
  • GenCorrect on non contest tasks. A published result applying the same loop to a software engineering benchmark with a verifiable oracle, such as a repository level test suite, within a year, would show whether the method leaves the contest hall.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source
  • Contest rules change. If the IOI or similar competitions adopt new rules on AI participation for 2027, the benchmark has acknowledged its own saturation.3 Tech Times 2026-09-05 1.2 million reasoning traces distilled from DeepSeek-V4-Flash for Nano and 477,642 for Ultra; run in Uzbekistan under International Technical Committee supervision; same judging platform and time limits; competition specific system; prior systems reached gold without beating the top human. Open source

The contest has been won. The open question is whether the recipe that won it works anywhere a grader does not exist.