Nemotron 3 Ultra (550B A55B)

Nemotron 3 Ultra is NVIDIA's frontier-scale open model with 550B total / 55B active parameters, built for agentic reasoning, long-context analysis, tool use, and high-stakes RAG. It uses a hybrid Latent Mixture-of-Experts (LatentMoE) architecture interleaving Mamba-2, MoE, and select Attention layers, with Multi-Token Prediction (MTP) for native speculative decoding, and is pre-trained on ~20T tokens with an NVFP4 recipe. Reasoning is configurable on/off (plus a medium-effort mode) via the chat template. It supports up to a 1M-token context and 10 languages (English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, Chinese). Released with open weights, training data, and recipes under the OpenMDW-1.1 license.