Nvidia shows open models for document search at GTC Berlin

At GTC Berlin, Nvidia showed new open weights models for document search, PDF files and driving of cars. Each person can get open weights models. The new models work with text and images in 34 languages. Nvidia tells that the 8 billion parameter version is first on the ViDoRe test. The models are for European buyers who want control of their AI.

How it works

The Nemotron Retriever models change text and images into number lists, or embeddings. A reranker model then puts the results in sequence, with the right results first.

Nemotron Parse changes PDF files into data. Alpamayo 2 Edge is a small model for driving.

The numbers

• The 8 billion parameter models are first in dense models on ViDoRe, at BF16 precision.
• Parse does about 10.5 pages in one second on one H100 chip, with the 1 billion parameter model.
• Bolt has 74 billion parameters and fits in 24 gigabytes of memory.

Nvidia gives all of these numbers. The source shows no test by a different group.

What is not known

Nvidia gives no test result for Alpamayo on a road. Nvidia usually gives the training data and recipes with the weights.

Insight: Free Nvidia models seem to pull buyers to Nvidia chips

Nvidia gives its models, data and recipes to all. Its models have versions for Blackwell chips, and the Parse speed is for an H100 chip. In Nvidia tests, a Rerank model cut agent cost by about 14%. This seems a move to pull buyers to Nvidia chips: if free models work in a good way on them, sovereign AI can follow. Leaders must check the chip limits of an open model before they select it.