NVIDIA releases open AI models for document-heavy work and local agents

NVIDIA has released open-weight AI models for two connected jobs: reading complex business documents and running agents on local systems. The flagship, Nemotron 3.5 Super VL, combines visual and language tasks in a 120-billion-parameter model. NVIDIA’s strongest performance claims come from customized versions trained for narrow assignments—not from the base model beating rivals across the board.

Super VL is designed to work across PDFs, charts, tables, images and webpages while following instructions and using tools. Combining those abilities can reduce handoffs between separate systems that read text and visual material.

The next step is post-training: adapting an existing model to a narrower job. NVIDIA says organizations can train Super VL around their own data, tools and policies, then deploy it on infrastructure that meets their security and data-location requirements. Its named projects include Scale AI’s work on vendor-credit finance and acquisition due diligence, and AfterQuery’s work on credit-agreement analysis for Hebbia.

NVIDIA’s chart covers five specialized evaluations, from document intelligence to medical insurance claims. Each customized model scores above the named Claude or GPT rival on its particular task. These are different trained versions, not one model winning every test. VentureBeat notes that NVIDIA did not supply task definitions, sample sizes or testing protocols.

NVIDIA also pitches Super VL as up to five times faster than a comparable open model. Its chart shows roughly 625 output tokens—pieces of generated text—per second, versus about 125 for Inkling Small at a similar quality score. NVIDIA says Artificial Analysis measured Super VL on four Blackwell B200 GPUs; the comparison uses 100,000 input tokens.

That speed comes at an Artificial Analysis Intelligence Index score of about 26, below several other open models in the chart. The index combines 10 evaluations; image inputs are assessed separately, so it does not capture the full document-reading pitch. Nor does output speed measure a finished agent job: reading inputs, calling tools, checking results and correcting errors also take time.

The smaller Bolt 74B-A9B targets research, coding and other multistep work on local systems. It has 74 billion parameters—the model’s internal learned values—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, without sending those workloads to a cloud-hosted model.

Bolt E4B takes that approach down to 4GB. NVIDIA schedules its release for October 23, targeting voice assistants, interactive characters and embedded interfaces. The company says it supports multi-turn conversations and tool use informed by an application’s current state.

Access and customization

• Super VL weights are on Hugging Face, with OpenRouter, ModelScope and NVIDIA NIM also listed as access routes.
• Bolt 74B-A9B is on Hugging Face, with support for llama.cpp, Ollama, LM Studio and vLLM.
• NVIDIA says its OpenMDW license permits commercial use, fine-tuning and redistribution. It is also releasing training recipes and datasets.

Developers can try a rate-limited NVIDIA-hosted API at no charge; partner-hosted services set their own prices. For its production offering, NVIDIA quotes an AI Enterprise software license of $4,500 per GPU per year. That is not the full deployment bill: GPU hardware, hosting and operating costs come on top.