In this article
  1. What happened
  2. Why it matters
  3. What to watch
  4. What to do

TL;DR: On October 21, 2026, NVIDIA announced five open retrieval models: Nemotron 3.5 Embed (8B, 1B and a 1B NVFP4 version), Nemotron 3.5 Rerank 8B and Nemotron 3.5 Parse (631M parameters). Embed and Rerank work with document page images, including charts and tables, as well as text. Parse turns scanned or image-only PDF pages into structured text. The model cards say all of them are “ready for commercial or non-commercial use” under the OpenMDW-1.1 license. If you already use Nemotron 3 Embed for text search, NVIDIA says text embeddings are expected to stay the same, so your existing text index can be reused.

What happened

NVIDIA announced the models in a post on its Hugging Face blog, calling them “new open, commercially available models for embedding, reranking, and document extraction.” The weights are on Hugging Face, and NVIDIA also offers them as NVIDIA NIM containers.

According to NVIDIA’s post and the model cards:

  • Embed 8B and 1B. Each turns a text query, or a document page as an image, text or both, into a vector for search. The 8B model outputs 4,096 dimensions and the 1B model 2,048. Both can be cut down to shorter vectors, such as 1,024 or 512, if you re-normalize them. Context is up to 32,768 tokens.
  • Text index reuse. The cards say the text part of each model was kept frozen during training, so it shares Nemotron 3 Embed’s text embedding space. NVIDIA says existing text embeddings and vector indexes can be reused. Image search is the new part.
  • Embed 1B NVFP4. A compressed version for vLLM. It needs a GPU with NVFP4 support, such as NVIDIA Blackwell. NVIDIA says it delivers 27.1% higher throughput than the 1B BF16 model on a GB200 while keeping 99.85% of its retrieval accuracy on the ViDoRe v3 benchmark.
  • Rerank 8B. Scores each query and document pair, text or image, to reorder the top results from the embedding search.
  • Parse. Reads document page images and returns text with layout classes, bounding boxes and reading order, and can extract tables and charts. NVIDIA reports 10.5 to 16.1 pages per second on one H100, depending on how many passes you run.
  • Where to run them. Besides Hugging Face and NIM, NVIDIA lists Baseten (Embed, Rerank, Parse), DeepInfra and FriendliAI (Embed 1B, Embed 8B, Rerank 8B) as hosted options.
  • Dates. The Embed and Rerank cards list a Hugging Face release date of October 20, 2026, and the Parse card October 17. NVIDIA announced the set on October 21.

Why it matters

A lot of small-business knowledge sits in PDFs, scans, slide decks and spreadsheets saved as images. Text-only search misses what’s in a chart or a scanned table. These models aim to make that content searchable with open weights you can run yourself or through a hosted API.

NVIDIA also argues that better retrieval cuts the cost of agents. In its own test, adding Rerank to Embed cut total agent tokens per task by about 14% with the 8B embedder and 13% with the 1B, while improving retrieval accuracy. That’s NVIDIA’s benchmark on its own pipeline, not a guarantee for your documents. The leaderboard rankings it cites (first on ViDoRe v3.1 for dense retrieval and reranking, first on ParseBench under 1 billion parameters) are as of October 20.

What to watch

  • Hardware. The cards say the models are built for NVIDIA GPUs. The BF16 1B card lists Ampere, Hopper and Blackwell. The NVFP4 version is Blackwell-only.
  • License. The models use the OpenMDW-1.1 license, not Apache or MIT, so read it before you ship. The Embed and Rerank cards add that the Mistral Ministral 3 base models they build on are under Apache 2.0, and the Parse card says its tokenizer is under the MIT license.
  • Re-indexing takes compute. Adding page images means running every page through the model again, which costs GPU time or hosted API fees. That’s our note, not NVIDIA’s.

What to do

  1. Start with text if you already have Nemotron 3 Embed. NVIDIA says your text vectors can be reused. Add page images only for documents where charts or scans carry the answer.
  2. Test on your own files first. Pick 20 to 50 real questions your team asks, run them against a small sample of your PDFs, and compare answers with and without Rerank before you re-index everything.
  3. Use Parse for scans with no text layer. If you only need fields out of a document rather than search, a hosted extraction model may be simpler. See our story on Cloudflare’s Clairic extraction model.
  4. Pick hosted or local on purpose. Hosted options avoid buying a GPU. Running locally keeps documents in-house. NVIDIA’s Nemotron 3.5 Bolt local agent model is from the same family if you want the whole stack on your own machine.
  5. Watch agent spend. If retrieval feeds an agent, set limits before you scale. Our guide to hard spending limits for AI agents covers how.

Sources & method

Sources

Method. Last checked October 21, 2026, at about 7:06 p.m. Manila time (11:06 UTC). Model names, sizes, vector sizes, context length, hardware, release dates, license terms and hosting partners come from NVIDIA’s post and the five model cards. Benchmark, speed and token-saving figures are NVIDIA’s own, and we haven’t tested the models. The advice in “What to do” is our own. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.