NVIDIA expanded its Nemotron family with a batch of new open-weight models - Nemotron 3.5 Super VL (a 120B multimodal model), Nemotron 3.5 Bolt (available in 74B and a 4B efficient variant), and Nemotron 3.5 Transcribe (a 30B mixture-of-experts speech model covering 40 languages). All are published on Hugging Face for agentic, enterprise, edge, and multilingual audio workloads.
Super VL: A 120B Multimodal Agent Model
Nemotron 3.5 Super VL is a 120-billion-parameter multimodal model built for specialized agent tasks. Reporting from VentureBeat and Interesting Engineering notes it outperforms rivals such as Claude Opus 5 and GPT-6.1 Sol on specific agent benchmarks and can complete tasks up to five times faster.
Bolt and Transcribe
Nemotron 3.5 Bolt ships in two sizes: a 74-billion-parameter version for heavier enterprise work and a 4-billion-parameter efficient variant (E4B) for edge devices and local deployment. Nemotron 3.5 Transcribe is a 30-billion-parameter mixture-of-experts model for transcription, translation, and speaker recognition across 40 languages, trained on a 463,062-hour audio dataset.
Key points
- 120 billion parameters - Super VL multimodal model
- 74 billion and 4 billion parameters - Bolt sizes
- 30 billion parameter MoE - Transcribe speech model
- 40 languages supported by Transcribe
- 463,062 hours - Transcribe audio training dataset
- Open weights available on Hugging Face
Read the full report on ThreeRouter AI Daily: https://www.threerouter.com/blog/415
