NVIDIA has released "Nemotron-3-Labs-Ultra-Math-RL," a decoder-only Transformer language model specialized in mathematical inference. This model is trained to solve complex mathematical problems and identify errors in proofs. It is operated as part of an ensemble system, and NVIDIA stated that it achieved gold-medal level scores in the 2026 International Mathematical Olympiad (IMO).

The model architecture employs Hybrid Latent Mixture of Experts (LatentMoE), combining Mamba2 and Transformer, and features a Multi-Token Prediction (MTP) network. It has a total of 550B parameters, with 55B active parameters. It supports a maximum context length of 1 million tokens.

This model was developed to improve mathematical inference using reinforcement learning. The dataset "Nemotron-Math-Proofs-v3-RL" was used for training mathematical proof generation. It is available for both commercial and non-commercial use.

Source: