AI Radar

Nemotron-4 340B

Available · Language

Nemotron-4 340B is a family of language models from NVIDIA (base, instruct and reward), released on 14 June 2024 as open weights on Hugging Face and the NGC catalog under the NVIDIA Open Model License. NVIDIA presented it for generating synthetic data to train other language models and sized it to fit on one eight-GPU DGX H100 in FP8 precision. [1] [2] [3] Primary source

Timeline of Nemotron-4 340B →

Claims and evidence

  • Released Primary source[1]page date; first and fourth paragraphs
  • NVIDIA-hosted API endpoint for Nemotron-4-340B-Instruct deprecated The notice gives 04/02/2025, read month first as NVIDIA writes dates elsewhere (its blogs use forms such as 10/21/2024). Applies to the hosted API in the NVIDIA API catalog only, not to the downloadable weights. Primary source[4]deprecation notice
  • Status AvailableWeights remain downloadable from NVIDIA's Hugging Face organisation. The NVIDIA-hosted API endpoint for the Instruct model was deprecated (see milestones). Primary source[2] [5]
    • Change · Size 340 billion parameters, against 15 billion for Nemotron-4 15B, with a similar architecture.Compared with Nemotron-4 15B Primary source[3]
    • Change · Training data Uses the same data blend as Nemotron-4 15B but trains on 9 trillion tokens: the first 8 trillion plus 1 trillion tokens of continued pre-training.Compared with Nemotron-4 15B Primary source[3]
    • Change · Availability Weights released for download under the NVIDIA Open Model License, whereas Nemotron-4 15B was published as a technical report only.Compared with Nemotron-4 15B Primary source[1] [2]
    • Open weights YesReleased under the NVIDIA Open Model License, which allows commercial use and distribution of derivative models. Primary source[1]fourth paragraph [2]License [3]Abstract
    • Context window 4,096 tokens tokens Primary source[2]Model Overview [5]Model Overview
    • Parameters 340 billionThe technical report gives 9.4 billion embedding and 331.6 billion non-embedding parameters. Primary source[2]Model Overview
    • Access Open-weights download, APIDownloads from NGC and Hugging Face at launch; the announcement said access at ai.nvidia.com as an NVIDIA NIM would follow, and the Instruct card points to build.nvidia.com. The hosted Instruct endpoint was later deprecated. Primary source[1]fourth paragraph [5]Model Overview [4]

    Lineage

    Predecessors

    No known predecessor.

    Successors

    No known successor.

    Based on

    Not derived from another model.

    Variants and derived

    None recorded.

    Siblings

    None recorded.

    All ancestors

    None.

    All descendants

    None.

    Variants

    Nemotron-4-340B-Base

    Same dates as Nemotron-4 340B.

    • Variant Inline variant in this record.Base completion model, pre-trained on 9 trillion tokens. Primary source[2]Model Overview [3]Abstract

    Also known as: nvidia/Nemotron-4-340B-Base

    Nemotron-4-340B-Instruct

    Same dates as Nemotron-4 340B.

    • Variant Inline variant in this record.Fine-tuned from the Base model for English single- and multi-turn chat; alignment used SFT, DPO and NVIDIA's Reward-aware Preference Optimization. Primary source[5]Model Overview [3]Abstract

    Also known as: nvidia/Nemotron-4-340B-Instruct

    Related AI Radar coverage

    AI Radar coverage starts in June 2026; no coverage linked yet.

    All model releases from NVIDIA on AI Radar →