🤖Nvidia Unveils Open Weights for Chatbots, Healthcare, and Robotics
New Nvidia models for RAG workflows in 34 languages
TL;DR
Nvidia released open weights models for RAG workflows, supporting embeddings and reranking for images and text in 34 languages. Models range from 1B to 8B parameters, with NVFP4-quantized versions for Blackwell-based accelerators.
Nvidia just dropped open weights models for chatbots, healthcare, and robotics, designed for retrieval augmented generation (RAG) workflows. These models support embeddings and reranking for images and text across 34 languages, with sizes ranging from 1B to 8B parameters. The 8B parameter model tops the ViDoRe V3.1 benchmarks at BF16-precision. If you're working on multimodal applications or need cross-language support, this is a big deal. The models can run on a DGX Spark with room to spare, and the Nemotron Parse model can process 10.5 pages per second from PDFs.
Key Points
Nvidia's models support embeddings and reranking for images and text in 34 languages.
Models range from 1 billion to 8 billion parameters, with NVFP4-quantized versions for Blackwell-based accelerators.
At BF16-precision, Nvidia's 8B parameter model tops the ViDoRe V3.1 benchmarks.
Nemotron Parse model processes 10.5 pages per second from PDFs on a single H100.
Alpamayo 2 Edge promises real-time processing of streaming video and sensor data on DRIVE AGX Thor.
Why It Matters
If you're working on multimodal applications or need cross-language support, Nvidia's new models offer a significant boost. The 8 billion parameter model leads the pack on benchmarks, and the Nemotron Parse can process PDFs at a high rate. This is a game changer for anyone looking to build sovereign AI solutions.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,569 builders reading daily.