NVIDIA has released Nemotron 3.5 Super VL, an open, customizable multimodal model aimed at enterprises building specialized AI agents. The model uses a 120-billion-parameter mixture-of-experts architecture that activates only 12 billion parameters per token, a design NVIDIA says balances frontier-scale capacity with inference efficiency. The weights are available on Hugging Face in BF16 precision, and NVIDIA published a technical walkthrough alongside the release.
Super VL slots into NVIDIA's expanding open-model family as the multimodal specialist. According to the company, smaller Nemotron 3.5 Lightning models handle high-volume text tasks, while the frontier Nemotron 3 Ultra tackles complex planning and orchestration — letting teams route each step of an agent workflow to the right size of model. It is the latest move in a push ZDNET described as another step forward for US open-weight AI. If you are tracking the open-weights race, our coverage of recent open model releases has the broader picture.
One Model for Text, Images, Documents and Video
The core pitch is consolidation. Super VL combines reasoning, instruction following and tool use with text, image, document and video understanding, so teams do not need to deploy separate text and vision-language models or shuttle task context between them. NVIDIA says its visual training covers charts, tables, optical character recognition, multipage documents and video, with a two-stage supervised fine-tuning pipeline that fuses visual evidence with reasoning, chat, safety and long-context data.
The target workflows are the ones enterprises keep stumbling over: a customer-service agent that must read company policy documents, a research agent connecting evidence across sources, or a document agent that interprets a chart and links it to a table elsewhere in the same filing. In each case the model needs to see the material and reason over it in the same context window.
Throughput Claims: Up to 5x on a Single GPU
Efficiency is where NVIDIA is drawing the sharpest lines. The company claims Nemotron 3.5 Super VL delivers up to 5x higher throughput for agentic tasks compared with other open models in its size category on a single GPU, and pointed to benchmarks showing more than 600 tokens per second on Artificial Analysis Intelligence Index agentic tasks via the Baseten inference endpoint.
Several techniques under the hood drive those numbers. NVIDIA highlighted LiLiCorr, an open speculative decoding method, along with NVFP4 mixed precision, multi-token prediction, and context reuse across agent turns — a meaningful optimization for agents that hold long-running conversations with tools. For enterprises running agents nonstop, throughput per GPU is often the difference between a viable product and an unusable cost structure.
Post-Training Is the Point
NVIDIA is explicit that the base model is a starting point rather than a finished product. Developers can customize Super VL with the NeMo library using supervised fine-tuning and LoRA to teach domain-specific behavior, then apply reinforcement learning through NeMo RL and NeMo Gym in environments that mirror their own operations. The company says organizations in cybersecurity, finance and healthcare are already post-training the model to beat frontier alternatives on their production workflows, a claim that is hard to verify independently but consistent with the enterprise-agent positioning.
To support that work, NVIDIA is releasing open datasets on Hugging Face alongside the weights, including agentic tool-use and conversational data, finance reasoning and tool-calling sets, instruction-following and chat data, long-context transduction and translation data, fairness data, and 40 vision-language datasets. Prime Intellect has already published a recipe for post-training the model for long-horizon enterprise work, an early sign the ecosystem is treating the release as a serious foundation. NVIDIA also published a usage cookbook on GitHub covering customization, and the company stresses that the open weights can be deployed in local, data center or cloud environments to meet infrastructure and data-residency requirements — the flexibility that regulated buyers consistently ask for and closed APIs cannot offer.
What It Means for the Open-Weight Race
The release tightens the gap between open and closed models in the segment enterprises actually deploy: mid-size multimodal systems for document-heavy agent work. Because the weights can run in local, data center or cloud environments, Super VL gives regulated industries an option that frontier APIs cannot — keeping sensitive documents on their own infrastructure while still getting chart, table and video understanding.
The competitive read is straightforward. Meta, Alibaba, DeepSeek and Liquid AI have all shipped open-weight models aimed at agents this year, and NVIDIA — which sells the hardware everyone else trains on — is now competing directly in the model layer with permissive licensing and an efficiency story. If the throughput claims hold up in independent benchmarks, the default stack for enterprise agents in 2027 may be considerably more open than it looked a year ago.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
