NVIDIA releases free search models that read charts and tables

NVIDIA has released a new set of open search models called Nemotron 3.5. They help AI assistants find the right information in documents, including charts, tables and infographics, not just plain text. The company says they are cleared for commercial use. The top model ranked first among its type on a public document-search test.

What it does

Five models make up the release. Nemotron 3.5 Embed comes in three versions: 8B BF16, 1B BF16 and 1B NVFP4. It turns documents into a form that is quick to search. Nemotron 3.5 Rerank 8B BF16 then re-sorts the results so the best evidence comes first. Nemotron 3.5 Parse, at 631M parameters, handles document extraction. The Embed model also reads document images, such as datasheets and schematics.

The numbers

On the ViDoRe v3.1 test, the 8B Embed model scored 61.17 nDCG@10, a standard score for how well the right results show up near the top. NVIDIA says that ranks first among dense retrieval models as of October 20, 2026. The previous NVIDIA model, Llama Nemotron Embed VL 1B, scored 52.23.

• The 1B Embed model scored 58.38, which is 95.4% of the 8B score.
• On MIRACL Vision, the 8B model scored 72.6, beating Qwen3-VL-Embedding-8B by 4.3 points.
• The 8B Rerank model scored 60.49 on the combined image-and-text test.

Why cost comes into it

NVIDIA's argument is about wasted effort. When an AI agent gets poor search results, it searches again and opens more documents. That uses more tokens, the units of text AI services count and often charge for. NVIDIA tested agents with different search setups while keeping the agent, tools and tasks the same. It measured task success and total tokens used. The article points to a chart for those results, and the numbers are not in the text available here.

Who it is for

These are building blocks for developers and companies that want an assistant to answer from internal reports, spreadsheets and manuals. Most business owners will meet them through software built on top. The smaller 1B version suggests a cheaper option for teams that want most of the quality at a smaller size.

What is not clear yet

Most of these results come from NVIDIA's own write-up, and the test scores are not the same as results on a real company's files. The material available does not give pricing, hardware needs or detailed results for the Parse model. The part about the reranker's leaderboard result is also cut off. NVIDIA and the MTEB team are launching a new reranking leaderboard for ViDoRe v3.1, which may give outside comparisons.