Is this page helpful?

Release Notes for NVIDIA NIM for Nemotron Parse#

This documentation contains the release notes for NVIDIA NIM for Nemotron Parse.

Release 2.4#

This is the first release of NVIDIA NIM for Nemotron Parse as a standalone NIM. Previous versions of the Nemotron Parse model were served by the NVIDIA NIM for Vision Language Models (VLMs).

Nemotron 3.5 Parse is a 650M parameter VLM that brings multilingual text and layout understanding into an optimized NIM, helping teams prepare complex documents for search, RAG, and downstream AI workflows.

Highlights#

  • Nemotron Parse is now a standalone NIM.

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

    • The container runs the model nvidia/nemotron-3.5-parse only.

  • Added a document-oriented API, POST /v1/parse, that replaces the OpenAI-compatible chat-completions API used by the VLM NIM. This is a breaking change for users who migrate from the VLM NIM. For details, refer to Migrate from the VLM NIM.

  • Each request returns extracted page content with Markdown formatting and coordinate/class control tokens. The NIM owns the prompt and generation settings, so requests don’t include prompts, special tokens, or sampling parameters. Detection-only requests are not supported.

  • Added the options.decoding.max_output_tokens request option, from 1 through 9,000 tokens, to limit the output length. The response reports finish_reason and usage.output_tokens.

  • Each request contains one rendered page as a PNG or JPEG image. To parse a PDF, render each page to an image first. For details, refer to Parse a PDF.

  • Added optional KServe V2 gRPC inference. For configuration and request examples, refer to API Reference (gRPC) for NVIDIA NIM for Nemotron Parse.

  • Added decoder-capacity controls, fail-fast admission, and diagnostic timing and VRAM logging. Refer to Environment Variables for NVIDIA NIM for Nemotron Parse.

Support Matrix#

  • BF16 (optimized) on NVIDIA GB200, NVIDIA RTX PRO 6000 Blackwell Server Edition, NVIDIA H100 NVL, NVIDIA A100, NVIDIA L40S, NVIDIA A10G.

  • BF16 is the only supported precision in this release.

For details, refer to Support Matrix.

All Known Issues#

The following are the known issues in release 2.4:

  • Dense-page output truncation — Especially dense pages may reach the maximum output limit of 9,000 tokens and return incomplete content. The response reports finish_reason: "length". The API has no continuation or resume option. As a workaround, crop the page into smaller regions and parse each region separately, then check the combined output for missing or repeated content.

  • Fixed prompt — The NIM does not permit prompt override; the supported prompt is fixed as </s><s><predict_bbox><predict_classes><output_markdown><predict_text_in_pic>. This differs from the Hugging Face model, which accepts other prompts.

For input limits, output truncation, and deployment troubleshooting, refer to Troubleshoot NVIDIA NIM for Nemotron Parse.

Release Notes for Previous Versions#

Previous versions of the Nemotron Parse model were served by the NVIDIA NIM for VLMs. For earlier release notes, refer to the following: