Hybrid-capacity MoE design
Uses a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token, balancing total model capacity with more selective computation.
NVIDIA Nemotron 3.5 Super VL
NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for specialized AI agents, combining reasoning, tool use, and NVIDIA NeMo support.
Overview
NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for specialized AI agents. It is designed for multistep enterprise workflows that may require reasoning, tool calls, policy following, and interpretation of text, images, documents, and video.
Its 120B-parameter mixture-of-experts architecture activates 12B parameters per token. The model can serve as the multimodal reasoning component in a broader Nemotron system, alongside smaller models for high-volume text tasks and larger models for complex planning and orchestration. Developers can adapt it with NVIDIA NeMo for domain-specific data, tools, and operating procedures.
Core capabilities
Multimodal agent reasoning
Combines reasoning, instruction following, and tool use with understanding of text, images, documents, and video in one model.
Visual and document understanding
Supports document workflows that interpret tables, charts, and page layouts, as well as video search, review, and incident analysis.
Workflow customization
Can be post-trained on an organization’s data, tool interfaces, operating procedures, and policies using NVIDIA NeMo.
Inference optimization
Uses LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse to improve inference efficiency.
Reported throughput
The launch article reports over 600 output tokens per second on an Artificial Analysis Intelligence Index agentic-task endpoint and up to 5x higher throughput than other open models in its size category on a single GPU; results depend on the stated benchmark and deployment conditions.
Practical use cases
Coding agents
Test and validate code, analyze errors, and recommend fixes in workflows that combine reasoning with tool use.
Web research
Search across websites, connect information from multiple sources, and verify answers before producing a result.
Document intelligence
Read tables, charts, and page layouts to extract information and generate summaries or reports from complex documents.
Cybersecurity analysis
Identify vulnerabilities and triage security threats so teams can prioritize remediation work.
Video understanding
Extract information from video for search, content review, and incident analysis.
Pros and Cons
Pros
- Combines reasoning, instruction following, tool use, and multimodal understanding in one model.
- Supports customization for domain data, tools, policies, and workflows through NVIDIA NeMo.
- Covers distinct agent tasks including coding, web research, document intelligence, cybersecurity, and video understanding.
- Uses a selective mixture-of-experts architecture and multiple inference optimizations intended for efficient agentic workloads.
Cons
- The supplied source does not provide a complete model card covering deployment requirements, supported platforms, outputs, licensing, or known limitations.
- Reported speed and throughput figures are benchmark-specific and should not be treated as universal performance guarantees.
- Pricing and distribution terms are not specified in the supplied evidence.
FAQ
What kinds of workflows is Nemotron 3.5 Super VL designed for?
Nemotron 3.5 Super VL is intended for specialized, multistep agent workflows that combine reasoning, tool use, and understanding of text, images, documents, or video. Examples include coding, web research, document intelligence, cybersecurity, and video analysis.
Can organizations customize the model?
The model can be post-trained with NVIDIA NeMo using organizational data, tools, policies, and workflows. The source also identifies supervised fine-tuning, low-rank adaptation, and reinforcement learning as customization methods, with open datasets released on Hugging Face.
What modalities does the model support?
The model combines reasoning, instruction following, tool use, and multimodal understanding in one model, so a workflow can process text and visual inputs without requiring separate text and vision-language models for those capabilities.
How is inference efficiency addressed?
The source describes deployment optimization through techniques including LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse. It points developers to NVIDIA NIM for optimized deployment, but the supplied evidence does not specify a complete deployment configuration or supported platform matrix.
Quick Facts
- Category
- Multimodal AI model
- Primary users
- Developers and organizations building specialized AI agents
- Architecture
- 120B-parameter mixture of experts; 12B parameters activated per token
- Input understanding
- Text, images, documents, and video
- Customization
- NVIDIA NeMo post-training, including supervised fine-tuning, low-rank adaptation, and reinforcement learning
- Source domain
- developer.nvidia.com
NVIDIA Nemotron 3.5 Super VL Alternatives
AakarDev AI
AakarDev AI helps teams manage AI provider access, project setup, logs, and analytics in one dashboard. BYOK support included.
Benchspan
Benchspan is an AI agent security platform that discovers agents, blocks prompt injection and data exfiltration in real time, and supports pre-launch red teaming.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
GitBot
GitBot is a local browser workspace for reusable coding bots powered by Claude Code, Codex, or OpenCode, with persistent threads and configurable permissions.
Codex Plugins
Codex Plugins bundle reusable skills, app integrations, and MCP servers into workflows you can install in the Codex app or use from Codex CLI.
Grok 4.7
Grok 4.7 is an xAI model for coding and knowledge work, available via Grok Build, Cursor, the Grok API and more, from $2 per million input tokens.