NVIDIA releases open source Nemotron 3.5 Lightning model for local AI

Thanks to the availability of open weights, developers and researchers can customize the system to adapt it to specific workflows, such as writing texts with predefined styles and formats, understanding vertical technical areas such as 3D graphics or computer programming and managing artificial intelligence agents used for home automation or individual productivity. To facilitate implementation on local systems, collaborations with platforms such as vLLM, Ollama, llama.cpp, LM Studio and Unsloth have been activated, ensuring native support for NVFP4 and GGUF quantized formats through Unsloth Studio.

Hardware compatibility ranges from personal computers equipped with RTX cards to NVIDIA Jetson industrial systems, expanding towards professional workstations, data centers and cloud infrastructures distributed by the main hardware partners. At the same time, she was released on GitHub the open source library NeMo Switchyard, a routing tool designed to analyze each operational phase of an autonomous agent and direct it towards the most suitable model based on accuracy, speed and cost per token criteria; From published internal benchmarks, the routing system allowed high-end performance levels to be maintained while reducing completion costs to approximately one-third compared to the isolated use of proprietary models such as Opus 4.8. Technical resources and application guides have been made available on the blogs dedicated to Nemotron 3.5 Lightning, NeMo Switchyard and the Jetson AI environment, as well as the distribution of the model via the build.nvidia.com API as a NIM microservice and on the OpenRouter platform.