14h agoNVIDIA details Nemotron 3 Ultra 550B pre-training using NVFP4, LatentMoE, and 20 trillion tokensNVFP4 training achieved a 0.4% relative loss gap vs BF16.SentimentSentimentPos100%Neg0%Many users praise NVIDIA's tech reports on Nemotron models and NVFP4 for their transparency and openness of pretraining data subsets.8 comments with sentiment. View comments.