NVIDIA Unveils New Nemotron Models: Super VL for Data Centers and Bolt for Edge Devices

Image: NVIDIA · Source

NVIDIA expands its Nemotron AI model family with the high-efficiency 120-billion-parameter Super VL for enterprises and two Bolt edge models optimized for local workloads, enabling customizable AI agents across computing tiers.

NVIDIA has broadened its Nemotron family of AI models with the launch of Nemotron 3.5 Super VL, a 120-billion-parameter open model designed for specialized always-on AI agents requiring fast task completion and high accuracy. Alongside Super VL, NVIDIA introduced two new edge-focused models: Nemotron 3.5 Bolt 74B-A9B 24GB, balancing intelligence and speed for local agentic workloads, and the smaller Bolt E4B 4GB, optimized for rapid conversational AI agents. While Super VL and Bolt 74B-A9B are immediately available, Bolt E4B is set to launch on October 23, 2026.

Super VL delivers frontier-level accuracy on complex multimodal tasks such as instruction following, tool utilization, business workflows, and visual document understanding, but with roughly one-fifth the parameters of NVIDIA’s largest reasoning model. It outpaces competing open models in speed by a factor of five, enabling enterprises to accelerate workflows that often extend over days or weeks. Unlike proprietary models, Super VL’s open weights allow organizations to post-train on proprietary data and deploy on their own infrastructure, ensuring control and privacy.

The Bolt series enables AI agent intelligence at the edge. Bolt 74B-A9B 24GB employs 2-bit quantization to fit within 24 GB memory, running efficiently on hardware like NVIDIA RTX and DGX Spark systems while supporting concurrent applications. This model is recognized for balancing speed and intelligence, ranking #1 on PinchBench among open models in its class. Bolt E4B 4GB targets conversational and physical AI agents on diverse hardware platforms, delivering multi-turn interaction capabilities for voice assistants and embedded AI.

NVIDIA highlights widespread adoption and customization of Super VL, with businesses across finance, cybersecurity, coding, scientific research, and healthcare post-training the model for domain-specific tasks. Partners such as Scale AI, EY, AI21, Atlassian, CrowdStrike, Salesforce, and others leverage Super VL for tasks ranging from due diligence to security and medical claims analysis.

Super VL’s multimodal vision-language capabilities enable it to comprehend text, tables, charts, images, and webpages fluidly, allowing AI agents to integrate visual and textual inputs seamlessly without task handoffs between systems. The model and datasets are fully open, shipped with training recipes, and are accessible via platforms including Hugging Face and NVIDIA NIM, which packages the model as a microservice optimized for enterprise deployment.

The Nemotron 3.5 models are compatible with a wide array of NVIDIA hardware including DGX Spark, DGX Station, RTX PRO, Jetson, and GeForce RTX PCs, bridging data center-scale AI and local edge intelligence. This launch further complements NVIDIA’s August release of Nemotron Lightning and Ultra models dedicated to high-throughput and advanced reasoning workloads.

Complementing these launches is the NVIDIA Nemotron Coalition, a global partnership of AI labs and developers including Mistral AI, LangChain, Perplexity, and others, collaborating on the development of open frontier models. The coalition’s first project, a model co-developed by NVIDIA and Mistral AI, will support the upcoming Nemotron 4 family, emphasizing transparency, collaboration, and AI sovereignty.

NVIDIA’s continued investment in open, customizable AI models aims to democratize advanced artificial intelligence while enabling organizations to maintain control over data and AI deployment across diverse operational environments.