NVIDIA Expands Nemotron 3.5 Lineup With High-Efficiency Agentic Models

NVIDIA has expanded its open-weight model portfolio on October 15, 2026, introducing the 120-billion-parameter Nemotron 3.5 Super VL alongside two specialized edge-optimized variants. This release underscores a strategic pivot toward agentic workflows, providing developers with high-efficiency tools designed for both enterprise-grade reasoning and localized, low-latency execution.

The flagship Nemotron 3.5 Super VL model functions as a multimodal engine for always-on agents, specifically engineered to handle complex business workflows. NVIDIA reports that the model achieves performance parity with its largest reasoning architectures while utilizing only one-fifth of the total parameter count. This efficiency gain is attributed to a combination of architectural refinements that prioritize instruction following and tool use.

Visual comprehension remains a core strength of the Super VL model, which features specialized capabilities for optical character recognition and structured data analysis. By integrating text and visual processing into a single pipeline, the model eliminates the need for task routing between disparate systems. This architecture allows for the direct extraction of insights from visual documents such as PDFs and web pages.

NVIDIA has also introduced the Nemotron 3.5 Bolt 74B-A9B 24GB, a hybrid mixture-of-experts model designed for local agentic workloads. Although the model contains 74 billion parameters, it activates only 9 billion per inference, utilizing 2-bit quantization to fit within a 24GB memory footprint. This configuration enables multistep reasoning and web browsing tasks to occur entirely on local hardware, enhancing privacy for sensitive enterprise data.

The third addition, the Nemotron 3.5 Bolt E4B 4GB, serves as a lightweight solution for conversational and physical agents. Optimized for platforms including GeForce RTX and Jetson hardware, this model focuses on rapid local voice interaction and application state grounding. It provides a low-latency interface for embedded systems that require multi-turn tool calling and precise instruction adherence.

All three models incorporate architectural elements from the previously released Nemotron 3.5 Lightning, including a hybrid mamba transformer and multi-token prediction. These design choices aim to address the latency bottlenecks often associated with large-scale agentic tasks. By utilizing speculative decoding and latent mixture-of-experts structures, the models attempt to balance high-volume throughput with the demands of real-time interaction.

The release includes comprehensive training recipes and datasets, with the Super VL model having been trained on over 25 trillion tokens. NVIDIA intends for these models to function as production-ready microservices via the NVIDIA NIM framework. By providing open weights, the company allows enterprises to perform post-training on proprietary data, ensuring that the models remain adaptable to specific organizational requirements.

The shift toward these specialized architectures reflects broader industry trends in MLOps, where the focus is moving away from raw parameter counts toward inference efficiency. By enabling local execution on hardware like the DGX Spark, NVIDIA is reducing the reliance on cloud-based API calls for everyday agentic processes. This decentralization of compute resources is expected to lower operational costs for high-volume enterprise deployments.

The performance of these models in real-world scenarios will likely depend on the effectiveness of the hybrid mamba transformer implementation. While the 120B Super VL model offers significant improvements in multimodal reasoning, the true test for the Bolt series will be its ability to maintain accuracy under the constraints of 2-bit quantization. Engineers should monitor how these models handle complex long-context tasks compared to traditional dense architectures.

Future deployments will likely center on the integration of these models into existing MLOps pipelines via support for vLLM and Ollama. The industry will be watching to see if the performance gains observed in the Super VL model translate into measurable reductions in latency for large-scale, multistep business automation. Continued development of these edge-optimized tools suggests a long-term commitment to balancing high-end reasoning with hardware-constrained execution environments.