Details
- NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model for always-on agents, alongside Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB edge models.
- The models target deployments ranging from enterprise servers and cloud infrastructure to RTX PCs, DGX Spark, DGX Station, Jetson devices and embedded systems.
- Super VL handles instruction following, tool use, multistep workflows and visual understanding across webpages, PDFs, tables and charts; Bolt 74B-A9B uses native 2-bit quantization to fit within 24GB and support local agent workloads.
- Super VL and Bolt 74B-A9B became available in October 2026, while Bolt E4B is scheduled for October 23. Bolt 74B-A9B supports llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
- NVIDIA says Super VL offers up to five-times the throughput of comparable models on Artificial Analysis agentic tasks, while its research documentation targets at least 80 tokens per second for Bolt 74B-A9B.[1][4]
Impact
The release strengthens NVIDIA’s strategy of pairing open model weights and customization tools with its hardware stack, potentially shifting more agent inference from cloud APIs toward controlled enterprise and local deployments. Over the next 12–24 months, its support for post-training, open datasets and widely used inference frameworks could expand developer adoption while increasing demand for RTX, DGX and Jetson systems.