NVIDIA rolls out Nemotron 3.5 Super VL plus new edge “Bolt” models to expand open agent deployments

NVIDIA on October 15 released Nemotron 3.5 Super VL, an open-weight artificial intelligence model designed for “always-on” agents that can complete tasks reliably and quickly. The company positioned the new model as its most efficient open option for multimodal agent work, then broadened the lineup with two additional open models built for edge and local deployment: Nemotron 3.5 Bolt 74B-A9B (24GB) and Nemotron 3.5 Bolt E4B (4GB).

Nemotron 3.5 Super VL is an open model with 120 billion parameters. NVIDIA said it delivers faster task completion and higher accuracy for specialized agents, and it targets common real-world agent scenarios such as instruction following, tool use, multi-step business workflows, and visual understanding. The company also said the model’s performance is comparable with NVIDIA’s largest reasoning model while using about one-fifth of the parameter count, and it claimed it is the fastest open agentic model in its class, at “5x faster than the next best,” with improvements aimed at high-volume workflows that can span days or weeks.

A central part of NVIDIA’s pitch is open customization. Unlike closed “frontier” models, the company said organizations can post-train Super VL on their own data, deploy it on their own infrastructure, and oversee what they build. NVIDIA described a workflow that supports post-training and customization through NVIDIA NeMo, Python-based pipelines, and open datasets, with the model shipping with open weights, open datasets, and training recipes.

For deployments that need visual and document understanding, NVIDIA emphasized Super VL’s vision-language strengths. The model is designed to process web pages by extracting text, reading tables, and interpreting graphs and charts, then returning insights from what it finds. NVIDIA also said the same approach applies to PDFs and other visual documents, allowing an agent to carry context from document review to downstream decisions without moving between separate systems.

NVIDIA also disclosed some training-scale details. It said Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples plus reinforcement learning across dozens of environments. To support production use, NVIDIA said it is packaging the model as NVIDIA NIM, which it describes as a production-ready microservice with performance-optimized configurations, enterprise-grade support, and security patches and performance updates refreshed every two weeks.

To widen the “data center to desktop” reach, NVIDIA introduced two Bolt variants aimed at smaller hardware footprints and local use cases. Nemotron 3.5 Bolt 74B-A9B (24GB) is designed for everyday agentic tasks locally on NVIDIA RTX and DGX Spark systems, using native 2-bit quantization to fit within 24GB of memory while running alongside other applications. NVIDIA said the model is “smart enough” to handle most tasks while still being much faster than larger options that would be harder to fit on typical local setups.

NVIDIA said Bolt 74B-A9B (24GB) can support delegating projects to popular local agent harnesses, including workflows such as web research, writing, and debugging code, without sending those workloads to the cloud. The company also cited benchmark positioning, stating it is the top open model in its class on PinchBench and the second open model for BrowseComp. It added that the model is available now on Hugging Face and that it will be supported for NVIDIA RTX Spark laptops arriving October 16.

The third release, Nemotron 3.5 Bolt E4B (4GB), targets even smaller local deployments. NVIDIA said it is built for physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark platforms, combining natural conversation with multi-turn agentic capabilities for fast local voice assistants, interactive characters, and embedded interfaces. NVIDIA said Bolt E4B is expected to arrive on October 23, and it described the Bolt family as covering the range of agent deployments from enterprise servers down to edge devices.

The broader market context is that companies are increasingly trying to deploy AI agents close to where data is created and decisions are made. NVIDIA’s framing suggests it wants to reduce the tradeoff that often appears in agent deployments: choosing between higher capability models and practical deployment constraints like latency, cost, and data residency. The company also indicated that multiple organizations across industries are already post-training Super VL for specialized tasks, including credit agreement analysis, M&A due diligence, cybersecurity attack resilience, and coding-agent benchmarks, among other use cases.

Still, the company’s release leaves some open questions for outside observers. NVIDIA’s benchmark and speed claims are presented at a high level, without publishing full evaluation methodology in the release text. It also did not specify typical throughput, latency figures, or total deployment costs for NIM services, nor did it provide external independent validation beyond internal positioning and cited benchmark standings. As with many open-model rollouts, security, reliability under adversarial inputs, and real-world integration timelines will likely depend on each organization’s post-training approach and system design.

Looking ahead, the key items to watch are adoption by developers and enterprises and how quickly integrators plug Super VL and Bolt into agent toolchains. NVIDIA said the Super VL model is already available on multiple platforms including Hugging Face, Open Router, ModelScope community, and as NVIDIA NIM, while Bolt E4B begins later in the month. Monitoring early production deployments, reported performance in long-running workflows, and how well the models sustain accuracy when tools and document understanding are combined could clarify the practical impact of NVIDIA’s expanded open Nemotron lineup.