NVIDIA releases open Nemotron 3.5 models for agent document search

NVIDIA released 5 open Nemotron 3.5 models for commercial use: 3 Embed, one Rerank and one Parse. They help AI agents find the right evidence in text, tables and charts. An embedding model changes content into numbers, which a search can compare. NVIDIA tells that better retrieval cuts the tokens (the text units that an agent pays for) which an agent uses.

How it works

Embed finds candidate documents, also from images of tables, charts and datasheets. Rerank puts the candidates in sequence for each query, and the agent sees the best evidence first. Parse changes documents into data that an agent can read.

The numbers

• Rerank cut agent token cost for each query by about 14% with Embed 8B and 13% with Embed 1B.
• On ViDoRe v3.1, a document search test, Embed 8B has a score of 61.17, first in dense retrieval models. The NVIDIA model before it has 52.23.
• Embed 1B keeps 95.4% of the 8B score.

What is not known

NVIDIA did the tests. The source names no external group that did them again. It does not tell how the results hold on other data.

Insight: NVIDIA fills each layer of the agent stack with open models

On 15 October, Nscale added the open Nemotron 3.5 Super VL model from NVIDIA. NVIDIA tells that it can operate on one Blackwell chip. NVIDIA adds retrieval models, and it started the ViDoRe v3.1 reranking leaderboard with the MTEB team. NVIDIA makes the models and helps to operate the scoreboard, which reads as a plan to set the standard for agent search. Test the models on your own documents.