NVIDIA Launches Nemotron 3.5 Super VL and Bolt Edge Models
New multimodal models expand agentic capabilities from enterprise servers to local edge hardware.
1New Multimodal Model Architecture
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, featuring a 120-billion-parameter mixture-of-experts architecture. The model activates 12 billion parameters per token, enabling high-accuracy multimodal tasks. It integrates reasoning, instruction following, and tool use with text, image, and video understanding, allowing teams to process complex visual documents like PDFs and charts efficiently.
2Performance and Benchmark Results
Super VL demonstrates significant gains, scoring 63.44% on Hermes Bench and 87.05% on LangChain evaluations. NVIDIA reports the model achieves up to 5x higher throughput than comparable open models on a single GPU. Independent evaluations by EXO Labs placed the model on the Pareto frontier for accuracy and task completion time.
3Collaborative Training and Distillation
The model was distilled from Nemotron 3 Ultra using multi-teacher on-policy distillation. Datasets from Perplexity supported mid-training, while Prime Intellect provided environments for teacher models. The training process utilized 26 trillion tokens, including 10 billion Indic tokens, with additional datasets developed in partnership with IIT Madras and the AI4Bharat lab.
4Edge Deployment with Bolt Models
NVIDIA introduced two Bolt models to target edge and local hardware. The 74B-A9B version uses native 2-bit quantization to fit within 24GB of memory, supporting local execution on RTX PCs and DGX Spark. The smaller E4B model, arriving October 23, 2026, is designed for embedded interfaces and voice assistants on Jetson devices.
5Optimized Serving and Quantization
The release includes NVFP4-quantized checkpoints using 4-bit weights for Blackwell architecture. Speculative decoding via LiLiCorr delivers 2.3x faster token generation, while multi-token prediction and context-reuse optimizations preserve attention caches. NVIDIA NIM packages these models as production-ready microservices, refreshed every two weeks to ensure ongoing security patches and performance improvements for enterprise users.
6Enterprise Agentic Workflow Integration
Major organizations are post-training Super VL for specific workflows. Scale AI customized the model for vendor credit finance, projecting a 40x reduction in inference costs. Meanwhile, Ramp expects 50% higher throughput for receipt processing. Other firms like Atlassian, CrowdStrike, and Salesforce are integrating the model to automate search, security, and government operations.
7Technical Capabilities for Developers
Developers can customize models using NVIDIA NeMo and Python-based pipelines. NVIDIA provides open datasets on Hugging Face covering finance, tool use, and 45 vision-language categories. The release supports various frameworks including llama.cpp, Ollama, and vLLM, ensuring compatibility across diverse hardware environments ranging from data centers to local workstations and laptops.
8Strategic Shift in Agentic AI
By covering the full deployment spectrum from large servers to edge devices, NVIDIA aims to standardize agentic AI infrastructure. The collaboration with the Nemotron Coalition and the release of open weights signal a move toward accessible, high-performance reasoning. This strategy enables businesses to deploy specialized agents that handle complex, multi-step workflows locally.
Original article: Unite.AI — NVIDIA Launches Nemotron 3.5 Super VL and Bolt Edge Models