NVIDIA announced on Thursday, October 15, 2026, the Nemotron 3.5 Super VL, an open 120-billion-parameter model designed for always-on AI agents that need to handle complex multimodal tasks efficiently.
The model uses a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture, activating about 12 billion parameters per token. It combines reasoning, instruction following, tool use and understanding of text, images, documents and video in a single system — eliminating the need for separate vision and language models for many workflows.
What changes in practice
According to NVIDIA, Super VL delivers the highest accuracy on multimodal agentic tasks in its class, with performance comparable to the company's largest reasoning model using roughly one-fifth the parameters. It is described as the fastest open agentic model in its category, up to 5x faster than the next best in throughput.
It processes PDFs, charts, tables, web pages and images, extracts information and acts on it without losing context. The context window reaches 1 million tokens. The model was trained on about 26 trillion tokens and post-trained with millions of samples and reinforcement learning.
Companies such as Scale AI, EY, Hebbia, AI21, Mercor, Perplexity, Salesforce and others are already customizing the model for specific cases, from financial document analysis to cybersecurity and coding agents.
Availability and hardware
Open weights are available now on Hugging Face (including optimized NVFP4 version) and via NVIDIA NIM. It runs on DGX Spark, DGX Station, RTX systems and data center environments.
Along with Super VL, NVIDIA launched the Bolt line for edge: the Bolt 74B-A9B (already covered by Geekiki, optimized for 24 GB on local hardware like RTX Spark) and the 4 GB Bolt E4B, arriving on October 23.
The Nemotron 3.5 family now covers from enterprise servers to local devices, with open models that organizations can post-train with their own data and policies.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult original sources for full context.
By GeekikiBot
