By Hamza Faisal
NVIDIA is making its biggest open-model push yet for the agentic AI era. On October 15, the chip giant released Nemotron 3.5 Super VL, an open 120-billion-parameter model built for specialized always-on agents — and paired it with two smaller Bolt models that bring agent intelligence down to desktops and edge devices.
Super VL and Nemotron 3.5 Bolt 74B-A9B are available now, with the tiny Bolt E4B 4GB arriving on October 23. Together, NVIDIA says, the three models cover the full range of agent deployments — from large enterprise servers to local hardware — without forcing anyone to choose between capability and deployability.
A 120B model built for always-on agents
Super VL is NVIDIA's fastest open agentic model in its class, according to the company. It is designed for multimodal work: reading PDFs, charts, tables and webpages, following instructions, using tools, and executing multistep business workflows. An agent reviewing a credit agreement, for example, might need to extract figures from tables, interpret a chart, check a related webpage and apply company rules — all before producing a useful answer.
NVIDIA says Super VL matches the accuracy of its largest reasoning model at roughly one-fifth the parameter count, and finishes complex workflows up to five times faster than comparable open models in its class. The weights are open: organizations can post-train it on their own data, tools and policies, then deploy it on their own infrastructure — behind their own firewall, under their own data-residency rules.
VentureBeat reports the post-trained Super VL beats Claude Opus 5 and GPT-6.1 Sol on specialized task benchmarks, and independent analysis from Artificial Analysis puts it at 4x the task-completion share of its predecessor at far lower token cost.
Bolt brings agents to the edge
The two Bolt models handle the other end of the spectrum. Bolt 74B-A9B 24GB targets local coding, research and multistep tasks, using 2-bit quantization to fit in 24 GB of GPU memory — runnable on RTX laptops and workstations. Bolt E4B 4GB, due October 23, targets fast conversational agents on ultra-light devices like Jetson and desktop RTX cards.
The lineup slots into NVIDIA's existing Nemotron family: Lightning, released in August, handles high-throughput workloads; Ultra tackles the hardest reasoning; and Bolt brings intelligence to the edge. All three new models can be downloaded from Hugging Face, OpenRouter and ModelScope, and run on NVIDIA's DGX and RTX hardware.
Open models, enterprise money
Adoption is already underway. NVIDIA lists dozens of companies that have fine-tuned Super VL on private data, including Scale AI for finance agents, AfterQuery for credit-agreement analysis, and TrainLoop for dermatology prescription drafting. Platform partners including Atlassian, CrowdStrike, Salesforce and Perplexity are integrating the model into their own agent products.
The training story is open too: NVIDIA says Super VL was trained on about 26 trillion tokens of open data across pretraining and continued pretraining, with 3.8 million post-training samples and reinforcement learning across 100 open environments. It is also releasing open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, safety and vision-language understanding.
The timing is deliberate. Enterprises are racing to put agents into production but want their data to stay inside the building — and NVIDIA is positioning open weights as the answer: rent nothing, own the model, run it where your policies say it runs.
By Hamza Faisal
