An enterprise agent reviewing a credit agreement may need to extract figures from tables, interpret a chart, check a related webpage and apply the company’s own rules before it can produce a useful answer.
NVIDIA’s latest Nemotron includes three models that target this set of tasks and more: a large model that combines language and visual understanding, alongside two smaller models designed for local agents.
Nemotron 3.5 Super VL is a 120-billion-parameter open model that NVIDIA says can work across PDFs, charts, tables, images and webpages while following instructions and using tools.
Nemotron 3.5 Bolt 74B-A9B 24GB targets local coding, research and other multistep tasks.
Bolt E4B 4GB, due to be released on October 23, is aimed at conversational and device-based agents.
The potential value for enterprises is the ability to adapt a model to their documents and procedures, then choose where it runs. NVIDIA says organizations can post-train Super VL on their own data, tools and policies and deploy it on infrastructure suited to their security and data-residency requirements. Bolt offers a smaller option for keeping certain agent workloads on local systems.
What changes within Nemotron
Super VL is not NVIDIA’s first model to read visual material.
Nemotron Nano 2 VL already offered document and image analysis.
April’s Nemotron 3 Nano Omni combined text, images, audio and video in a 30-billion-parameter model that NVIDIA pitched for document intelligence and reading graphical interfaces.
The change is where visual understanding sits in the broader lineup. NVIDIA released a 120-billion-parameter Nemotron 3 Super model for agentic reasoning in March, followed by the larger Nemotron 3 Ultra for more demanding work. August’s Nemotron 3.5 Lightning targets the repeated, high-volume steps in longer agent workflows. Super VL brings document and image understanding into the larger Super tier, while Bolt extends the family’s local deployment options.
That combination could simplify some agent designs. A system reviewing an insurance claim, for example, might need to read scanned records, extract structured information, check it against a policy and decide what to flag for a person. Handling the text and visual material in one model can reduce the need to pass information between separate reading systems. Developers still have to test whether the model extracts the right facts and follows the relevant rules.
NVIDIA does not have the field of open multimodal models to itself. Meta’s Llama 4 and Google’s Gemma 4 also handle visual and textual inputs. Super VL’s particular case rests on the larger Nemotron model, NVIDIA’s tools for adapting and deploying it, and the results partners report after training it for specialized work.
What partners report after customization
NVIDIA says dozens of organizations are post-training Super VL for their own tasks.
Its announcement names work by Scale AI for Skydance’s vendor-credit finance agent and EY’s merger-and-acquisition due-diligence agent; AfterQuery for Hebbia’s credit-agreement analysis; AI21 and Alice for cybersecurity; and Mercor for coding agents. Ramp, Dream, Lila Sciences and R1 are among others customizing the model.
Platform companies integrating it include Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys.
An NVIDIA chart illustrates why those partners might customize rather than simply deploy the base model. It shows five separate evaluations: document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims.
The light-green portion of each bar represents base Super VL; the dark-green portion shows the reported gain after post-training. The improvements range from 9 to 44% across the five tasks.
In each comparison shown, the resulting specialized model scores above the named Claude or GPT model tested on that task:
The medical-claims result is the largest improvement over the base model: the chart shows accuracy rising from roughly 28% to 72% after post-training, compared with about 38% for Claude Opus 5 on that evaluation.
Security attack resilience rises from roughly 35% to 76%, compared with about 66% for GPT-6.1 Sol.
These are different customized versions on different specialized tasks, not a claim that the base Super VL model beats those models generally.
NVIDIA has not supplied the task definitions, sample sizes or testing protocol with the chart, so a buyer should test the same approach against its own work.
Fast output at a particular quality level
NVIDIA also describes Super VL as up to five times faster than a comparable open model. Its speed chart plots the Artificial Analysis Intelligence Index against output tokens per second with 100,000 input tokens.
Super VL appears near 625 output tokens per second at an index score of about 26. The arrow marking the fivefold comparison begins at Inkling Small, near 125 tokens per second and a similar index score.
NVIDIA says Artificial Analysis measured Super VL’s 625-token-per-second result on four Blackwell B200 GPUs.
The result makes Super VL unusually fast at that score in the displayed comparison. It is not the highest-scoring model overall: the accompanying Artificial Analysis chart places it at 26 on the composite index, below several other open models, including GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40.
The index combines 10 evaluations across areas including professional work, coding, science and reasoning. Artificial Analysis evaluates image inputs separately, so this composite score alone does not measure Super VL’s full document-reading pitch.
Output speed can matter when an agent makes many calls over a long job, but it is distinct from the time needed to complete that job. Reading inputs, calling outside tools, checking results and correcting mistakes all contribute to the total. The chart establishes a generation-speed comparison under its stated test conditions; it does not show a fivefold reduction in end-to-end workflow time.
Bolt brings agent work to smaller systems
Bolt 74B-A9B has 74 billion total parameters but activates 9 billion at a time. NVIDIA says native 2-bit compression lets it fit in 24GB of memory and run alongside other applications on RTX Spark and DGX Spark systems.
It is intended for work such as web research, writing and debugging code through local agent tools, without sending those workloads to a cloud-hosted model.
NVIDIA calls Bolt the top open model in its class on PinchBench and the second-ranked open model on BrowseComp, although its announcement does not give the scores or define the comparison class.
The model is available on Hugging Face with support for llama.cpp, Ollama, LM Studio and vLLM. NVIDIA also lists GeForce RTX PCs, DGX Station and RTX PRO workstations among supported systems.
Bolt E4B takes the local approach down to 4GB for voice assistants, interactive characters and embedded interfaces. NVIDIA says it supports conversations across multiple turns and can use tools while responding to an application’s current state. It targets GeForce RTX, Jetson and RTX/DGX Spark when it arrives October 23.
Where developers can get the models, and what production costs
Super VL’s open weights are available on Hugging Face. NVIDIA also lists OpenRouter, ModelScope and NVIDIA NIM as launch access routes. Bolt 74B-A9B is available on Hugging Face.
NVIDIA says the models use its OpenMDW license, which permits commercial use, fine-tuning and redistribution. It is releasing training recipes and datasets as well, and provides customization resources through NeMo and a technical architecture guide.
Developers can use a rate-limited NVIDIA-hosted API at no charge for development, while NVIDIA partners offer hosted APIs at prices set by each provider, including through OpenRouter.
Organizations can download NIM microservices from NVIDIA’s NGC catalog. For its production offering, NVIDIA quotes an NVIDIA AI Enterprise license of $4,500 per GPU per year. That is a software-license figure, not the total cost of the GPUs, hosting and operations needed to run Super VL.
NVIDIA says Super VL is optimized for Blackwell and Hopper systems, including DGX Spark, DGX Station and RTX PRO 6000 Blackwell GPUs, and can also run on Ampere-generation GPUs.
Its stated training totals are approximately 26 trillion tokens across pretraining and continued pretraining, 3.8 million post-training samples and reinforcement learning across roughly 100 open environments. NVIDIA also plans datasets covering tool use, finance, long context, safety and vision-language understanding.
The launch follows NVIDIA’s agreement to acquire Hugging Face, a major distribution platform for open models. The roughly $12.93 billion deal has not closed: NVIDIA expects completion in the first half of 2027, subject to regulatory approval and other conditions. The company has committed to keeping Hugging Face open to other model makers and chip vendors.
For enterprise developers, the reason to try Super VL is specific: test whether a model they can adapt and deploy under their own control reads their documents and follows their procedures accurately enough to justify its cost. Bolt offers a separate test for work they want to keep local. NVIDIA now provides the models, access routes and a production software price to begin those evaluations; the partner examples and speed chart show what to investigate, rather than what every deployment will achieve.