Nemotron-70B

Released October 15, 2024, Llama-3.1-Nemotron-70B-Instruct is NVIDIA's alignment-tuned variant of Meta's Llama 3.1 70B, post-trained with NVIDIA's RLHF and DPO pipeline including the then-#1 RewardBench reward model. It keeps the 128K context window and scored strongly on instruction-following and alignment benchmarks, and NVIDIA unusually published much of the alignment recipe — datasets, RL environments, and NeMo training code — alongside the weights.

The weights are downloadable under the NVIDIA Open Model License, which permits commercial use but is not OSI-open: it restricts training competing foundation models on the outputs and carries field-of-use conditions, so verify the terms before commercial deployment. As a 2024-era non-reasoning model it has been superseded by NVIDIA's own Nemotron 3 reasoning family, but it remains a solid, well-documented open-weight 70B baseline for self-hosted chat and RAG.