NVIDIA's Nemotron-3-Ultra-CC outscores every human at IOI 2026
The 550B open-weight model tops the best coder by 37 points; a 30B sibling scores IOI gold. Plus: AISLE finds 6 curl CVEs big-lab tools missed.
Copy markdown
First AI to top every human at IOI
Nemotron-3-Ultra-CC (550B total, 55B active) scored 535.4/600 on the IOI 2026 problem set, beating the top human competitor's 498.27 by 37 points — NVIDIA says it's the first time any AI has outscored the highest-scoring contestant on an IOI set. Gold started at 361.
A 30B model that already scores gold
The smaller Nemotron-3-Nano-CC (30B total, 3B active) hit 468 at IOI 2025, clearing the 438 gold threshold — competition-grade coding at a size you can run on a single node. The base Nemotron-3 weights are already open under OpenMDW.
GenCorrect is the trick, not just the model
The scores come from GenCorrect: generate up to 1,000 candidate solutions, cluster to ~10 diverse ones, run them, and refine on execution feedback over five rounds. That lifted Ultra-CC from 344 to 502 on IOI 2025. It's a sample-cluster-execute loop you can wrap around almost any coding model today.
What's open, and what's held back
NVIDIA plans to release the Ultra-CC checkpoint plus runnable inference and eval recipes in its NeMo-Skills repo. The full training corpus stays closed over third-party redistribution limits, but the post-training procedure is documented in the paper.
Elsewhere: AISLE finds 6 curl CVEs the big labs missed
AISLE's agents filed 29 reports against curl; six became CVEs, patched in 8.22.0 — including a 25-year-old bug reported as the oldest ever. OpenAI's Codex Security and Anthropic's Mythos had each found zero on the same codebase.
Elsewhere: fable51-worlds turns maps into explorable 3D
PhiloLabs open-sourced fable51-worlds, where Claude Fable 5.1 agents build walkable 3D reconstructions from OpenStreetMap and USGS data — a Union Square scene with 129 storefronts, a 2.3km Kyoto route. Clone the repo and run one this weekend.