NVIDIA Releases Five Nemotron 3.5 Retrieval Models

An AI agent is only as good as the evidence it can find. When the documents it pulls back are incomplete or off-topic, it searches again, reads more and burns through tokens before it finishes the job. NVIDIA’s latest release is aimed squarely at that problem. On 21 October it added five “open, commercially available models for embedding, reranking, and document extraction” to its Nemotron Retriever collection, it said in a post on Hugging Face.

The models are Nemotron 3.5 Embed in three versions (an 8-billion-parameter model and two 1-billion-parameter models, one compressed to NVIDIA’s NVFP4 number format), Nemotron 3.5 Rerank at 8 billion parameters, and Nemotron 3.5 Parse at 631 million.

What each model does

The three types of model handle different stages of the same job.

Parsing comes first. Much company knowledge sits in PDFs and scans, where pulling out the raw text can scramble the reading order, split table values from their headers or lose charts entirely. Nemotron 3.5 Parse turns page images into structured text, tables and blocks, keeping track of where each element sat on the page, so the content can be indexed and searched.

Embedding makes that content searchable. An embedding model turns each passage or page into a list of numbers, a vector, that captures its meaning, so a query can be matched against millions of documents quickly. NVIDIA says Nemotron 3.5 Embed works on images of pages as well as text, including tables, charts, infographics and schematics.

Reranking is the second, closer look. Once a search returns a shortlist, the reranker scores each candidate against the question directly, pushing the evidence that actually answers it to the top. For an agent with limited room in its context window, that means less distracting material to read.

NVIDIA’s numbers

All the results below are NVIDIA’s own, and TSN has not seen independent tests.

On the ViDoRe v3.1 visual document retrieval benchmark, NVIDIA says Embed 8B scores 61.17 on a ranking measure called nDCG@10, “ranking first among dense retrieval models” on the leaderboard as of 20 October. That is 8.94 points above NVIDIA’s earlier Llama Nemotron Embed VL 1B. The smaller 1B version, distilled from the 8B, keeps 95.4% of that score, NVIDIA says.

For agents, the headline claim is cost. In NVIDIA’s tests, adding the reranker cut the agent’s token use per query by about 14% with Embed 8B and 13% with Embed 1B, while also improving retrieval accuracy. On a new ViDoRe v3.1 reranking leaderboard, launched with the MTEB benchmark team, NVIDIA puts Rerank 8B at 60.49 against 59.17 for Qwen3-VL-Reranker-8B.

NVIDIA says Parse ranks first on ParseBench among models under a billion parameters, scoring 75.20 while handling 10.5 pages a second on a single H100 GPU. A faster single-pass mode processes 16.1 pages a second at a lower score of 72.74. The compressed Embed 1B model, it says, delivers 27.1% higher throughput on its GB200 hardware while keeping 99.85% of its accuracy.

Who is trying it

NVIDIA also listed early partner work. Elastic customers can already reach the models through Elasticsearch, it says, while Perplexity, Vectorize and Turbopuffer evaluated them on their own benchmarks. Those results, too, come via NVIDIA’s post.

One practical detail may matter to existing users. NVIDIA says the text side of the new embedding models was frozen during training, so text embeddings should stay the same as those from its earlier Nemotron 3 Embed models, letting teams keep their existing search indexes.

The weights are on Hugging Face and through NVIDIA’s NIM service, with hosted access via Baseten, DeepInfra, FriendliAI and OpenRouter.