NVIDIA has introduced Nemotron 3.5 Lightning, an open and customizable model designed for running agentic tasks locally on PCs with NVIDIA RTX GPUs and DGX Spark systems. The model works in conjunction with NeMo Switchyard, an open-source routing library designed to balance accuracy, speed, and cost across different architectures.
Nemotron 3.5 Lightning delivers up to four times faster token generation and a reduction in execution time compared to comparable models in the 30% class. The system enables local refinement through custom examples for email, calendars, home automation devices, and programming assistants running on local code bases.
The company has also introduced solutions for scaling DGX Spark systems. Using NVIDIA Sync Cluster Assistant, two or more units can be joined into a high-speed cluster with workload routing and network health monitoring. The new NVIDIA Sync Resource Monitor provides historical and real-time analytics of CPU and GPU utilization for individual systems or entire clusters. DGX Spark also integrates the native version of Google Chrome for Linux ARM64.
Among the company's hardware-optimized open-weight models is Meta's Muse Glimmer, a 30 billion-parameter model designed for local agents that can exceed 200 tokens per second on the RTX 5090 graphics card. For local generative video creation and editing, the LTX-2.5 model offers a 2x performance boost and a 40% reduction in memory usage on RTX GPUs, DGX Spark, and DGX Station. The ecosystem of accelerated open models also includes Cosmos 3 Edge, MiniMax-H3, Poolside's Laguna S 2.1, DeepSeek-V4-Flash, Thinking Machines Lab's Inkling-Small, Unsloth Desktop, and Wan-Animate-2.