Nano: 8B (8,030,261,248); Super: 49B (49,867,145,216); dense, not MoE
131,072 tokens (128K), both variants
NVIDIA Open Model License, with underlying Llama 3.1 Community License (Nano) / Llama 3.3 Community License (Super) and their Acceptable Use Policies; commercial use subject to attribution, redistribution, naming and >700M-MAU additional-license conditions (not Apache 2.0)
Text-only reasoning and chat family comprising Llama-3.1-Nemotron-Nano-8B-v1 and Llama-3.3-Nemotron-Super-49B-v1. Nano is a dense decoder-only Transformer post-trained from Llama 3.1 8B Instruct for local RTX/edge use. Super compresses Llama 3.3 70B Instruct using distillation-based neural architecture search, with 31 of 80 blocks skipping attention and variable-width feed-forward blocks; it remains dense, not a Mamba hybrid or mixture of experts. Super received 40B knowledge-distillation tokens from FineWeb, Buzz-V1.2 and Dolma. Both use supervised fine-tuning for math, code, reasoning and tool calling, followed by RLOO and online reward-aware preference optimization, with final SFT/RPO checkpoint merging. A system prompt switches detailed thinking on or off for reasoning versus lower-overhead chat, RAG and tool use. NVIDIA reports reasoning-on MATH500 scores of 95.4% for Nano and 96.6% for Super, and AIME25 scores of 47.1% and 58.4%, respectively, evaluated at 32K sequence length. Reasoning is task-dependent: Nano strict-instruction IFEval falls from 82.1% with reasoning off to 79.3% on, and thinking can still occur when disabled. Primarily intended for English and code, with seven additional supported languages; base pretraining knowledge cutoff is 2023. Super targets an accuracy/throughput tradeoff, but the released BF16 checkpoint lists two 80GB GPUs for inference rather than guaranteeing single-GPU deployment. Ultra is not part of this launch-day weight release.