NVIDIA Open Sources NemoTron 3.5 Lightning: A Tooling Call Runner Specifically for Agents

According to OmniVision Beating monitoring, NVIDIA has released the open-source Nemotron 3.5 Lightning model, designed specifically for long-running Agents to perform "execute chores." The model has a total of 300 billion parameters, with only 30 billion activated per token, mainly responsible for tool invocation, result checking, and sub-Agent scheduling for high-frequency tasks.

NVIDIA's approach is to further specialize Agent division of labor. Complex planning is assigned to large models like Nemotron 3 Ultra, while repetitive execution is handled by Lightning. The company claims that its output speed can reach up to four times that of similar models. On PinchBench, Lightning achieves an accuracy of 86% and, under similar accuracy conditions, completes 10,000 tasks 30% faster than Qwen3.6 35B.

The model adopts an MoE architecture, specifically trained for Agent Harness, and incorporates multi-token prediction and inference decoding. Officially providing BF16 and NVFP4 weights, it can run on local devices such as RTX 5090, DGX Spark, and also supports llama.cpp, Ollama, LM Studio, and Unsloth.

It also emphasizes its customizability. NVIDIA has opened up weights, some training data, and training recipes, allowing enterprises to continue post-training for tasks related to code, security, law, etc. In the showcased cases by the company, Lightning has shown significant score increases across various professional tasks after post-training.