NVIDIA drops Nemotron 3 Ultra after IMO gold

Achieving IMO gold-medal performance through pure natural-language reasoning—without leaning on formal proof assistants like Lean—demonstrates that raw LLM mathematical reasoning has evolved far faster than expected. While previous high-profile benchmarks relied heavily on formal theorem proving pipelines, Nemotron 3 Ultra demonstrates rigorous Olympiad-level deduction using standard mathematical prose. Scoring 30/42 without internet access, code execution, or external calculators highlights immense gains in internal chain-of-thought and test-time reasoning. Providing open checkpoints offers researchers and developers immediate access to analyze and build upon gold-medal-grade reasoning architectures.