NVIDIA launched Nemotron 3.5 Super VL, expanding its family of open models for always-on AI agents. The new model covers data centers to local PCs with RTX Spark chips.
What is Nemotron 3.5 Super VL?
Nemotron 3.5 Super VL is a multimodal 120-billion-parameter model (around 12 billion active) optimized for enterprise agent tasks. It delivers accuracy similar to Nemotron 3 Ultra in instruction following and tool use, with a focus on speed. On post-trained specific tasks — such as document intelligence, accounting, scientific experimentation, insurance, and attack resilience — NVIDIA claims that Nemotron 3.5 Super VL outperforms closed models like Claude Opus 5 and GPT-6.1 Sol.
Alongside Nemotron 3.5 Super VL come the lighter Bolt models: Nemotron 3.5 Bolt 74B-A9B 24GB (MoE with 9B active and native 2-bit quantization for 24 GB) and Bolt E4B 4GB (scheduled for October 23). These Bolt models run locally on RTX Spark and DGX Spark systems, enabling agents without sending data to the cloud.
Why Nemotron 3.5 Super VL matters
NVIDIA RTX Spark PCs are already available via partners like Microsoft and Dell. With Nemotron 3.5 Super VL and the Bolts, companies can customize and run agents on-premises or locally. Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce adopt or test Super VL. The weights of Nemotron 3.5 Super VL are on Hugging Face, with support for llama.cpp, Ollama, LM Studio, and vLLM.
This offering reinforces NVIDIA’s focus on open infrastructure optimized for its chips, while frontier labs diversify hardware.
Sources: TechCrunch, NVIDIA official blog.
By GeekikiBot