NVIDIA has released Nemotron 3.5 Super VL, a 120-parameter open multimodal agent model designed for always-on enterprise tasks. Built as a mixture-of-experts architecture that activates 12B parameters per token, the new model processes text, images, documents, and video while handling reasoning, tool calls, and code generation within a single checkpoint. Alongside the flagship release, NVIDIA introduced two edge models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B, expanding deployment options for developers.
The model interleaves Mamba-2, mixture-of-experts, and attention layers, utilizing a C-RADIO vision encoder and a specialized connector to handle visual inputs. Trained on roughly 26 trillion tokens and distilled from Nemotron 3 Ultra, Super VL achieves up to 5x higher single-throughput performance compared to open peers through techniques like multi-token prediction, LiLiCorr speculative drafting, NVFP4 quantization-aware distillation, and context reuse. Developers can access the open-weights model via Hugging Face, OpenRouter, and TensorRT-LLM runtimes, with early enterprise applications already demonstrating significant cost and latency reductions.
This release provides enterprises with a powerful, unified multimodal tool that streamlines agentic workflows and reduces infrastructure complexity.