Nvidia Releases Open-Source Nemotron 3.5 Models Tailored for Agentic AI and Edge Hardware

Nvidia has expanded its lineup of open-weight artificial intelligence systems by releasing three new models in its Nemotron family: the flagship Nemotron 3.5 Super VL, along with the Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB.

The launch follows the release of Nemotron 3.5 Lightning in August and represents the company's first model rollouts since announcing an agreement to acquire the open-source machine learning hub Hugging Face last month, where all three models are now hosted.

Targeting Instant Completion and Local Agents

Leading the new release is Nemotron 3.5 Super VL, a 120-billion-parameter model designed for "always on" autonomous agents. According to Nvidia, the model provides fast-track completion, allowing it to execute tasks nearly instantaneously while maintaining high accuracy.

Alongside the 120B parameter model, Nvidia introduced two variants under the Bolt label:

• Nemotron 3.5 Bolt 74B-A9B 24GB: A mid-sized model optimized to deliver balance and processing speed.
• Nemotron 3.5 Bolt E4B 4GB: A lightweight model engineered specifically for conversational agents.

In contrast to partially open releases common across the industry, Nvidia has made all three models fully open, providing public access to their weights, training datasets, and underlying training recipes.

Countering Rivals and Supporting the GPU Ecosystem

The new models arrive as U.S. developers seek strong, open alternatives to compete with low-cost, high-performance open-weight models from China, such as Alibaba's Qwen family. According to Alexander Harrowell, an analyst at Omdia, developers are already shifting toward the new release, actively substituting models like Alibaba's Qwen 3.6 35B with Nemotron Bolt.

The move also serves Nvidia’s broader hardware interests. Major cloud providers and key customers are increasingly designing proprietary silicon—including Google with its TPUs, Amazon with its Trainium accelerator chips, and OpenAI with its in-development Jalapeño chip. By releasing optimized open models, alongside steps such as open-sourcing its Run.ai GPU orchestration platform, Nvidia aims to sustain demand for its own compute stack.

"They're doing this to support the wider ecosystem around their GPUs," Harrowell noted, pointing to an industry-wide push where chipmakers closely integrate models, software, and hardware.

Advancing Agent Workloads to On-Device Computing

Beyond raw benchmark performance, the architectural design of Nemotron 3.5 Super VL and the Bolt models features improved memory scaling, which expands the context window required for complex, multistep agentic workflows.

The compact footprints of smaller variants like the E4B also align with an industry push to move agentic execution out of centralized data centers and onto local machines. Systems like Nvidia DGX Spark and Apple Mac Minis provide the potential to run agent workloads locally, though Harrowell pointed out that broader adoption will require enterprises to adjust back to managing packaged software hosted directly on physical devices rather than relying strictly on cloud-hosted APIs.