NVIDIA has developed Nemotron-3-Labs-Ultra-Math-SFT, a decoder-only Transformer language model specialized in mathematical problem solving and identifying errors in proofs. The model is trained to solve challenging mathematical problems and pinpoint mistakes within proofs.

As part of an ensemble system, this model has a track record of achieving gold-medal-level scores in the 2026 International Mathematical Olympiad. Its architecture employs "Nemotron Hybrid LatentMoE," a hybrid structure combining Mamba2 and Transformer, and features a Multi-Token Prediction (MTP) network. It has a total of 550B parameters, with 55B active.

For training, a dataset for mathematical inference was used, including traces of proof generation, refinement, verification, and meta-verification. The model is available for both commercial and non-commercial use.

Source: