NVIDIA Open-Sources Nemotron 3.5 Lightning for AI Agents
NVIDIA releases open-source Nemotron 3.5 Lightning, a 30B parameter model with 3B activated for fast AI agents. 4x faster, 86% accuracy, fine-tunable. Also open-sources NeMo Switchyard.
NVIDIA has released the open-source Nemotron 3.5 Lightning model for continuously running AI agents.
It is a model with 30 billion parameters, of which only 3 billion are activated. Thanks to this, generation speed is up to four times higher than models of comparable size.
In PinchBench, it achieved 86% accuracy and completed 10,000 tasks 35% faster than Qwen3.6 35B with comparable accuracy.
The model can be fine-tuned on custom data, tools, and workflows. It is compact enough to run even on DGX Spark.
Along with it, NVIDIA NeMo Switchyard — a library that distributes different stages of agent work across models. Complex reasoning and planning can be delegated to advanced models, while mass task execution goes to the fast Lightning.
Comments
Weekly digest
The best of vibe coding, AI agents and open source — once a week, no spam.
Related articles
All articlesCerebras Kimi K2.6 Hits 981 tok/s, Beating Top GPUs in Inference
Cerebras launches 1T-parameter Kimi K2.6 on Wafer-Scale Engine 3, reaching 981 tokens/sec—6.7× faster than top cloud GPUs and 23× market average.
NVIDIA Open-Sources LocateAnything-3B for Faster Visual Localization
LocateAnything-3B uses parallel decoding to predict bounding boxes at once, improving detection in dense scenes and supporting UI, OCR and documents.
NVIDIA Nemotron 3 Ultra: 550B MoE Open-Weights Model for Agents
NVIDIA unveils Nemotron 3 Ultra, a 550B hybrid Mamba+Transformer MoE with open weights, faster inference, lower cost, and long-horizon agent planning.