nemotron-super3

NVIDIA Nemotron 3 Super is a large language model featuring LatentMoE and MTP architectures. This skill retrieves cited facts explaining these structural components and their operational mechanisms directly from the model card and research paper.

FP8 and NVFP4 are quantization formats supported in Nemotron 3 Super. This skill differentiates their precision levels and deployment characteristics by grounding answers in the model's evaluation and quantization documentation.

The Nemotron 3 Super training pipeline includes pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL). This skill outlines what occurs in each phase by routing through recipe summaries and paper markdown sections.

Recipe mapping connects Nemotron 3 Super open-source reproduction recipes to original research claims. This skill clarifies how recipes align with the paper by explicitly separating reproducible open-source components from reported research benchmarks.

No, this skill does not build training jobs or run evaluation benchmarks. It functions strictly as a knowledge desk, retrieving cited facts about Nemotron 3 Super architecture, training stages, and quantization from existing documentation.