NVIDIA's Nemotron-3 Ultra Math RL ships, part of an ensemble that hit IMO 2026 gold

NVIDIA has released Nemotron-3-Labs-Ultra-Math-RL, the flagship 550-billion-parameter member of its Nemotron-3 family, specialized for mathematical reasoning. The model card, published on Hugging Face, states the model is "specialized for mathematical reasoning, trained to solve difficult mathematical problems and identify mistakes in proofs, and deployed as part of an ensemble system that achieved a gold-medal level score at the International Mathematical Olympiad 2026."

What's new

Nemotron-3-Labs-Ultra-Math-RL uses a "Mamba2-Transformer Hybrid Latent Mixture of Experts (LatentMoE) with Multi-Token Prediction (MTP)" architecture, with 550B total parameters and 55B active per forward pass. It is built on the base model nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, and NVIDIA trained it on a proof-focused reinforcement-learning dataset, Nemotron-Math-Proofs-v3-RL, made up of 9,597 proof-generation prompts.

Key specs from the model card:

• 550B total parameters, 55B active (Mixture-of-Experts)
• Context length up to 1M tokens
• Minimum deployment footprint: 8x B200 GPUs on a single node
• License: OpenMDW-1.1
• Published to Hugging Face September 3, 2026, as part of the "Nemotron Labs IMO 2026" collection

The headline result is that the model, run as part of an ensemble system rather than standalone, reached a gold-medal-level score on the 2026 International Mathematical Olympiad — matching the bar cleared this year by other frontier labs' IMO efforts, but doing so with an open-weight release rather than a closed research demo.

Context

NVIDIA introduced the broader Nemotron 3 family with a 30B Nano variant available immediately and teased a 100B "Super" and a 550B "Ultra" tier as coming later. Nemotron-3-Labs-Ultra-Math-RL is that Ultra tier, now shipped and specialized specifically for math and proof-checking rather than as a general-purpose flagship. NVIDIA has increasingly used Nemotron releases to demonstrate specific reasoning capabilities rather than only raw scale, following a broader industry pattern of labs publishing IMO-caliber results as a proxy for frontier reasoning ability.

Why it matters

An open-weight model that contributes to an IMO-gold-level ensemble result is a different proposition than a closed frontier chatbot doing the same: researchers can inspect the architecture, reproduce the training recipe against the released proof dataset, and build on the checkpoint directly. The 1M-token context window and Mixture-of-Experts design also make it a case study in how far hybrid Mamba-Transformer architectures can be pushed for structured, verification-heavy reasoning tasks like proof-checking, rather than pure language generation. For NVIDIA, it converts a benchmark flex into a released artifact developers can actually run — provided they have the eight B200 GPUs the model card requires.