Is this page helpful?

Release Notes for NVIDIA NeMo Retriever Embedding NIM#

This documentation contains the release notes for NVIDIA NeMo Retriever Embedding NIM.

Note

Some releases are labelled “Production Branch” or “(PB)”. Production Branches provide reliable, stable versions of the NIM. Non-production branch releases (sometimes called Feature Branch (FB) releases) contain the latest features, improvements, and optimizations.

Release 2.4#

Release 2.4 provides one Nemotron embedding experience for enterprise documents consisting of text, page images, tables, charts, and infographics - with 1B efficiency and an 8B quality path behind the same NIM API.

Highlights#

  • Added the Nemotron 3.5 Embed 1B multimodal embedding model with two checkpoints: nvidia/Nemotron-3.5-Embed-1B-BF16 and nvidia/Nemotron-3.5-Embed-1B-NVFP4.

  • The nvidia/Nemotron-3.5-Embed-1B-BF16 checkpoint runs on NVIDIA GB200, NVIDIA B300, NVIDIA RTX PRO 6000 Blackwell Server Edition, NVIDIA H100, NVIDIA H100 NVL, NVIDIA A100, NVIDIA L40S, NVIDIA A10G, and NVIDIA GB10 (DGX Spark). For details, refer to Support Matrix for NVIDIA NeMo Retriever Embedding NIM.

  • The nvidia/Nemotron-3.5-Embed-1B-NVFP4 checkpoint runs on NVIDIA GB200, NVIDIA B300, and NVIDIA RTX PRO 6000 Blackwell Server Edition, where it is the default. For details, refer to Support Matrix for NVIDIA NeMo Retriever Embedding NIM.

  • Added the Nemotron 3.5 Embed 8B multimodal embedding model with one checkpoint, nvidia/Nemotron-3.5-Embed-8B-BF16.

  • Starting with release 2.4, the NIM container name and the model name are different, and the container can run different supported models.

    • The container is nvcr.io/nim/nvidia/nemotron-embed.

    • The 2.4 release container can run nvidia/Nemotron-3.5-Embed-1B-BF16, nvidia/Nemotron-3.5-Embed-1B-NVFP4, or nvidia/Nemotron-3.5-Embed-8B-BF16.

    • Use the model ID that /v1/models returns in the model field of API requests. For details, refer to Get Started.

  • Added Multi-Instance GPU (MIG) support for Nemotron 3.5 Embed 1B with the default NIM configuration. For details, refer to Multi-Instance GPU (MIG).

For details, refer to Support Matrix.

Embedding Output Support#

Nemotron 3.5 Embed supports three embedding output types in release 2.4: float, int8, and uint8. Set embedding_type to one of these values or omit it to use the default, float. The binary and ubinary output types are not supported in release 2.4.

Both encoding_format: "float" and encoding_format: "base64" are supported independently of the selected embedding type. Output quantization is separate from model-weight and inference precision. The three output types are supported with both BF16 and NVFP4 profiles. Other models retain their listed embedding-type support. For float and integer request examples and scoring guidance, refer to How to Specify Embedding Type.

All Known Issues#

The known issues for NeMo Retriever Embedding NIM are the following:

Release Notes for Previous Versions#