Llama Nemotron (Nano/Super)Open Weight

The Llama Nemotron family, launched March 18, 2025 at NVIDIA's GTC conference, was the company's entry into the reasoning-model wave that DeepSeek-R1 had set off that January. Nano (8B, distilled from Llama 3.1) and Super (49B, compressed from Llama 3.3 70B by neural architecture search) shipped first as open weights, with an Ultra tier announced for multi-GPU servers. Each model carried a switchable reasoning mode — step-by-step thinking toggled on or off through the system prompt, the same hybrid idea Claude 3.7 Sonnet had introduced three weeks earlier.

NVIDIA positioned the family explicitly as infrastructure for agentic AI platforms rather than as a chatbot, claiming up to 20% accuracy gains over the Llama bases and large inference-speed multiples from the compression work. The releases came with most of the post-training data and the tools to reproduce the recipe, and the Nano/Super/Ultra size ladder became the naming scheme the Nemotron line kept from then on.