NVIDIA Bolsters Edge AI with Nemotron 3.5 Bolt: A Lightweight Model for Local Agent Workloads

NVIDIA has expanded its Nemotron 3.5 family with the introduction of Nemotron 3.5 Bolt, an open vision-language model specifically engineered for edge devices. Released on October 15, 2026, this model is designed to facilitate always-on and multimodal agent workloads directly on hardware like DGX Spark, H200, or GB200. It features a hybrid Mamba-2 / Transformer MoE architecture with 120 billion total parameters (12 billion active) and a C-RADIO vision encoder, capable of processing text, images, and video (up to 2 minutes) to produce text outputs. The model supports up to a 1 million-token context. This release is particularly significant for developers and organizations focused on deploying AI at the edge. The ability to run advanced vision-language models locally addresses critical challenges related to data privacy, latency, and operational costs. For industries such as manufacturing, healthcare, and autonomous systems, where real-time decision-making and data sovereignty are paramount, Nemotron 3.5 Bolt offers a compelling solution. It allows for the creation of sophisticated AI agents that can operate independently of cloud infrastructure, reducing reliance on internet connectivity and mitigating data transfer risks. The introduction of Nemotron 3.5 Bolt aligns with a broader trend in the AI landscape towards more efficient, specialized models and the decentralization of AI processing. While large, general-purpose models continue to push the boundaries of AI capabilities in the cloud, there's a growing recognition of the need for performant, smaller models that can function effectively in resource-constrained environments. This trend is also evident in the increasing availability of open-weight models from various providers, fostering innovation and allowing for greater customization and integration into proprietary systems. In practice, this means practitioners should evaluate Nemotron 3.5 Bolt for applications where data must remain on-device or where network latency is a critical bottleneck. The open-weight release, along with accompanying datasets, training recipes, and NeMo Gym RL environments, provides a robust foundation for development. Developers can leverage this to build custom AI agents for tasks like real-time anomaly detection, local content moderation, or intelligent robotic control. The model's strong performance on benchmarks like GPQA Diamond, SWE-bench Verified, and OSWorld-Verified suggests it is well-suited for complex reasoning and coding tasks, even in an edge context. This move by NVIDIA underscores the ongoing shift towards making powerful AI more accessible and deployable across a wider range of hardware and use cases.