NVIDIA expanded its open model families to cover vision-language, retrieval, tabular data, biology, robotics, and quantum computing. Nemotron 3.5 Super VL and Nemotron 3.5 Bolt are new language models for agents. Nemotron Retriever adds embedding, reranking, and document parsing models. Kumo models and libraries target structured business data for forecasting and fraud detection. Cosmos 3.5 world models accelerate robotics and autonomous vehicle training. Alpamayo 2 Edge enables real-time reasoning for autonomous vehicles. BioNeMo RFD4-Proteina advances protein design with 13x speed gains over RFdiffusion3. Ising models support quantum computing calibration and error correction. All models are available on Hugging Face, GitHub, build.nvidia.com, or partner platforms. Nemotron 3.5 Bolt E4B 4GB launches on October 23.
What changed
Previous NVIDIA open model families lacked dedicated vision-language, retrieval, tabular, biology, robotics, or quantum computing models.
What it unlocks
Building and deploying domain-specific AI agents and applications using open models tailored to local languages and industries.
- 13x faster than RFdiffusion3
- 200 amino acid proteins
- Nemotron 3.5 Bolt E4B 4GB available Oct 23
What you need to act on it
- access to Hugging Face, ModelScope, OpenRouter, build.nvidia.com, or partner platforms
- NVIDIA NIM microservice deployment for production inference
Sources