NVIDIA released three new open-weight models on October 15, 2026, targeting a broad spectrum of agentic AI applications ranging from enterprise automation to local edge computing. The release includes the 120-billion-parameter Nemotron 3.5 Super VL, the 74-billion-parameter Bolt 74B-A9B, and the compact 4-billion-parameter Bolt E4B.
The flagship Nemotron 3.5 Super VL is engineered specifically for always-on agentic systems. This model integrates multimodal capabilities to handle instruction following, complex tool usage, and visual data analysis within a single architecture.
NVIDIA reports that the Super VL model achieves performance parity with its largest previous reasoning models while utilizing only one-fifth of the parameter count. The architecture is optimized for high-volume enterprise tasks, processing structured data from PDFs, tables, and web pages without requiring task routing between disparate systems.
To support deployment, NVIDIA is providing open weights, training recipes, and datasets derived from a training corpus exceeding 25 trillion tokens. The model is distributed as a production-ready microservice via NVIDIA NIM and maintains compatibility across the company’s hardware stack, including DGX Spark, DGX Station, and Jetson platforms.
For edge-based workloads, the company introduced the Nemotron 3.5 Bolt 74B-A9B 24GB, a hybrid mixture-of-experts model. This variant features 74 billion parameters but activates only 9 billion at any given time, utilizing 2-bit quantization to fit within a 24GB memory footprint.
The Bolt 74B-A9B model enables local execution of multistep reasoning, web browsing, and code debugging tasks. By keeping these processes on local hardware like the RTX Spark, users maintain data privacy without offloading sensitive information to cloud environments.
The third addition, the Nemotron 3.5 Bolt E4B 4GB, focuses on low-latency conversational and physical agents. This model is engineered for rapid response times in embedded interfaces and interactive voice assistants, running efficiently on GeForce RTX and Jetson hardware.
These new models incorporate architectural advancements from the earlier Nemotron 3.5 Lightning release, including a hybrid mamba transformer and multi-token prediction. The hybrid mamba transformer architecture allows the model to maintain state-space efficiency while processing sequences, effectively bridging the gap between traditional transformer attention mechanisms and linear-time sequence modeling. By integrating multi-token prediction, the models can generate multiple future tokens in a single forward pass, which significantly reduces the latency associated with sequential decoding in agentic workflows.
The integration of latent MoE structures and speculative decoding further refines this efficiency. Speculative decoding allows the model to use a smaller, faster draft model to propose token sequences, which are then verified by the larger model, ensuring that high token generation rates translate into proportional gains in overall agentic task speed. This combination of techniques addresses previous bottlenecks where high-speed token output failed to yield meaningful improvements in complex, multi-step reasoning tasks.
The shift toward these specialized, open-weight models signals a strategic pivot for NVIDIA as it seeks to dominate the agentic AI infrastructure market. By providing tools for enterprises to post-train these models on proprietary data, the company is positioning its hardware as the primary foundation for private, high-performance agentic deployments.
This release highlights the growing industry focus on balancing model intelligence with hardware constraints. As agents become more integrated into business workflows, the ability to perform complex reasoning on local or edge hardware—rather than relying solely on massive cloud clusters—becomes a critical differentiator for enterprise adoption.
The reliance on 2-bit quantization and hybrid architectures suggests that NVIDIA is prioritizing efficiency to maximize the utility of existing GPU memory capacities. This approach allows developers to deploy sophisticated reasoning agents in environments where power and memory are strictly limited, expanding the reach of agentic AI beyond server-side applications.
Future developments will likely center on how these models perform in real-world, high-concurrency environments compared to proprietary closed-source alternatives. Stakeholders should monitor the adoption rates across the vLLM, Ollama, and llama.cpp ecosystems, as these platforms will serve as the primary testing grounds for the practical viability of the new Bolt-class models.