Nvidia Releases Nemotron 3.5 Super VL Open Model for Continuous Enterprise AI Agents
Nvidia launched Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model using Mixture-of-Experts architecture with 12 billion active parameters, designed for background agent tasks and document analysis.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Nvidia released Nemotron 3.5 Super VL on October 15, 2026, an open-source multimodal model with 120 billion parameters designed for specialized AI agents that run continuously in the background, processing text, images, documents, and videos in enterprise settings. The model utilizes a Mixture-of-Experts architecture where only 12 billion parameters are active per token.
According to The Decoder, Super VL was distilled from Nemotron 3 Ultra, Nvidia's largest reasoning model. Nvidia claims that for tasks involving instruction following, tool calling, and multi-step business processes, Super VL matches Ultra's accuracy with approximately one-fifth the parameters. In agent tests from Hermes and LangChain, it reportedly scores higher than Ultra and beats its predecessor Nemotron 3 Super on coding benchmarks SWE-bench and Terminal-Bench.
Alongside the main model, Nvidia introduced two compact Bolt variants for local hardware: Nemotron 3.5 Bolt 74B-A9B 24GB for systems with at least 24 GB of memory, and the upcoming Bolt E4B 4GB for voice assistants. These models are heavily compressed to offer what Nvidia calls the 'highest intelligence per gigabyte of memory.'
All models are optimized for Nvidia hardware, including Blackwell GPUs and DGX Spark workstations. Nvidia positions this as enabling companies to fine-tune models with their own data and run them on private servers, avoiding third-party terms of service. However, performance metrics remain attributed to vendor self-testing without widespread independent reproduction.