NVIDIA released three distinct offerings simultaneously: Nemotron 3.5 Super VL, an open-weight model carrying 120 billion parameters with the ability to process both written text and images, alongside two lightweight Bolt models designed to run agentic workloads directly on a device.
With this release, software developers working inside enterprise environments gain access to a model that can be further trained on internal documents and company-specific workflows, then deployed across infrastructure aligned with their organisation's data-residency and security policies. Equipped to process PDFs, charts, tables, images, and webpages, the system carries out instructions and operates across a range of built-in tools.
Two product lines occupy different ends of the spectrum: the 74B-A9B variant, carrying 24GB of capacity, handles on-device coding and research workloads, whereas the E4B at 4GB — scheduled to arrive on October 23 — is built around conversational applications and agent-based deployment.
Visual processing capability in AI models predates Super VL's arrival at NVIDIA. Before Nemotron 3 Nano Omni — a 30-billion-parameter model unifying text, images, audio, and video within one architecture — debuted in April, a pair of predecessor releases had already addressed document and image analysis, the same ground Nemotron Nano 2 VL had covered. Two NVIDIA language models built for agentic reasoning — the 120-billion-parameter Nemotron 3 Super and the larger Nemotron 3 Ultra — were rolled out in succession beginning that March.
Built by August, the Nemotron 3.5 Lightning model is designed to handle the repetitive, high-frequency tasks that appear throughout extended agent workflows. Two additions arrived under the platform's expanded offering: Super VL, handling both document and image comprehension, and Bolt, which widens the options available for running the software on local infrastructure.
When an automated tool processes an insurance claim, it may pull data from scanned documents, organise that information into structured form, cross-reference it against the relevant policy terms, and flag anything that warrants closer attention. When a single model takes on both text and images, the flow of data no longer needs to hop across separate reading pipelines. The accuracy of fact extraction and rule compliance remains an open question for those building with the model.
According to the chipmaker, well over two dozen external groups have taken the Super VL model and adapted it through post-training for their own specific applications. Among the use cases highlighted in the announcement are a vendor-credit finance tool built by Scale AI for Skydance, a merger-and-acquisition due-diligence tool developed for EY, credit-agreement analysis delivered through AfterQuery on behalf of Hebbia, and cybersecurity applications drawing on work from both AI21 and Alice.
Among the platform companies incorporating the technology into their offerings are Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, and Synopsys. Five task areas appear in an NVIDIA graphic: pulling structured data from documents, testing resistance to security attacks, extracting details from financial receipts, supporting scientific experimentation workflows, and processing medical insurance claims.
Measured against five separate tasks, performance after post-training climbed by anywhere from 9 to 44 percent. Across every paired comparison presented, the fine-tuned specialist model outperformed whichever Claude or GPT variant was put to the test on that particular task. On the medical-claims evaluation, post-training lifted the model's accuracy from around 28% to 72% — a swing of roughly 44 percentage points, while Claude Opus 5 topped out near 38%. Measured against a baseline near 35%, attack-resilience climbs to 76% — a level that outpaces the roughly 66% recorded for GPT-6.1 Sol.
Each variant has been fine-tuned for a distinct task, meaning no assertion is being made that the base Super VL model outperforms those alternatives across the board. No task definitions, sample sizes, or testing protocol have been made available by NVIDIA.
According to NVIDIA's own characterisation, the Super VL model runs at throughput roughly five times that of a comparable open-source alternative. A visualisation within the chart maps the Artificial Analysis Intelligence Index along one axis against token generation rate — measured at 100,000 input tokens — along the other. Running on a quartet of Blackwell B200 GPUs, Super VL registers a throughput of roughly 625 tokens per second alongside an index score of approximately 26. At the leftmost end of the five-model comparison, Inkling Small registers roughly 125 tokens per second alongside an index score in that same range.
Among the open-architecture entries on the index, GLM-5.3-Flash posted a composite tally of 42 while Qwen3.8-Flash-Next landed at 40, placing both ahead of several proprietary alternatives. The benchmark in that chart tracks output speed under the test conditions the tool itself defined — a boundary narrow enough to leave any assertion about a fivefold reduction in overall workflow time without support.
Of Bolt 74B-A9B's 74 billion parameters, only 9 billion are called into use during any given operation. A built-in 2-bit compression scheme keeps the model's footprint within 24 GB of RAM, leaving enough headroom for other applications to run at the same time on RTX Spark and DGX Spark platforms. On the capability side, web research, code composition, and debugging are handled through on-device agent tools. Deployment options for the model span four inference frameworks — llama.cpp, Ollama, LM Studio, and vLLM — with the weights hosted on Hugging Face.
Designed to run on GeForce RTX, Jetson, and RTX/DGX Spark hardware, the Bolt E4B model trims its memory footprint down to 4GB, opening it up for use in voice assistants, interactive characters, and embedded interfaces, with availability set for October 23.
Developers can access Super VL's open weights through Hugging Face, while OpenRouter, ModelScope, and NVIDIA NIM each serve as separate distribution channels for the initial rollout. The Bolt 74B-A9B model is available through the Hugging Face platform. Released under NVIDIA's OpenMDW licence, these models carry terms that allow developers to deploy them commercially, adapt them through fine-tuning, and pass them along to others.
Training materials — including recipes, datasets, and a technical architecture guide — are made available by NVIDIA alongside customisation support through the NeMo framework. Free-tier access to an NVIDIA-hosted API gives builders a starting point for development work, though throughput is capped by rate limits. Hosted APIs built on NVIDIA's partner network carry pricing determined independently by whichever provider delivers them — a structure that extends to platforms such as OpenRouter. NIM microservices housed in NVIDIA's NGC catalog are available for download by any interested organisation.
Deploying NVIDIA AI Enterprise in a production environment carries a listed software-licensing charge of $4,500 annually per GPU — a figure that covers only the licence itself, leaving out the underlying hardware, hosting infrastructure, and day-to-day operational expenses.
A deal placing Hugging Face under NVIDIA ownership — valued at around $12.93 billion — had already been struck by the time the product rolled out. Finalisation of the transaction remains pending, with NVIDIA placing anticipated closure sometime in the opening months of 2027, contingent on regulators granting their approval. Pledges to leave Hugging Face accessible to outside model developers and chip suppliers were put forward by the firm.
