Nemotron-4 340B

What it is

Nemotron-4 340B is a family of open models released by NVIDIA on 14 June 2024. The family has three members: Nemotron-4-340B-Base, Nemotron-4-340B-Instruct and Nemotron-4-340B-Reward ^[1][2]. A base model is trained on general text. An instruct model is further trained to follow requests. A reward model is trained to score how good a response is, which can guide the training of other models.

The family was aimed at generating synthetic data. Synthetic data is text or other material produced by a model rather than collected from people, and it can be used to train and refine other language models, including for commercial use ^[1]. In this sense the family was presented as a tool for building other models, not only as a model for people to chat with.

Origins

The line of Nemotron models began before this release. The predecessor in the line was Nemotron-3 8B, released on 15 November 2023 ^[3]. Nemotron-4 340B was the larger successor in the same line. Later, the Llama Nemotron reasoning models in 2025 continued the line ^[3].

Key features

The base model was trained on 9 trillion tokens ^[1]. A token is a small piece of text, often a word or part of one, and a larger training set usually gives a model more to learn from. It was sized to fit on a single DGX H100 with eight GPUs, using FP8 precision ^[2]. FP8 is an 8-bit floating-point format that stores each number in less memory than the usual 16- or 32-bit formats, which lowers the cost of running the model.

Alignment is the stage in which a model is adjusted to behave the way its developers want. Over 98% of the data used for alignment in this family was synthetically generated, and NVIDIA open-sourced the synthetic data pipeline that produced it ^[2]. Publishing the pipeline means that others can see how the data was made and reuse the method.

The reward model, Nemotron-4 340B Reward, took first place on AI2's RewardBench leaderboard on Hugging Face ^[1]. RewardBench is a benchmark for reward models, and a leaderboard ranks models on it.

The models were released under the NVIDIA Open Model License Agreement. That license allows distribution, modification and use of the models and their outputs ^[2]. The models were available for download from NVIDIA NGC and Hugging Face, and an NVIDIA NIM microservice was planned ^[1]. A NIM is a packaged service that makes a model easier to deploy.

The technical report, with Bo Adler and 80 other authors, was posted on 17 June 2024 ^[2]. It came three days after the release.

Where it sits

The family is part of NVIDIA's open model line, which later included the Llama Nemotron reasoning models in 2025 ^[3]. For students and developers, it is a clear example of the idea that a large model can be used to produce training material for smaller models, and that the method used to produce that material can be published alongside the models.