NVIDIA Nemotron 3.5 Lightning Open Source Release Targets Agent Tool Invocation

NVIDIA launches 30B-parameter MoE model Nemotron 3.5 Lightning for agent tool use, offering 4x speed and open weights for local fine-tuning.

Woofun AI reports that NVIDIA has released the open-source Nemotron 3.5 Lightning model, engineered specifically for execution tasks in long-running Agents. The 30-billion-parameter Mixture-of-Experts architecture activates only 3 billion parameters per token, optimizing performance for high-frequency operations like tool invocation and sub-Agent scheduling.

The model achieves output speeds up to four times faster than comparable models and completes 10,000 tasks 30% faster than Qwen3.6 35B while maintaining an 86% accuracy rate on PinchBench. NVIDIA provides BF16 and NVFP4 weight formats compatible with RTX 5090 and DGX Spark hardware, alongside support for llama.cpp, Ollama, LM Studio, and Unsloth. Training data and configurations are publicly available, enabling further customization for coding, security, and legal applications.