NVIDIA open-sources Nemotron 3.5 Lightning for agent tool execution

According to Beating Monitor, NVIDIA has released the open-source model Nemotron 3.5 Lightning, specifically designed to handle repetitive execution tasks for long-running agents. The model has a total of 30 billion parameters, with only 3 billion activated per token, primarily responsible for high-frequency tasks such as tool invocation, result verification, and sub-agent scheduling. NVIDIA’s approach is to more thoroughly specialize agent roles: complex planning is delegated to larger models like Nemotron 3 Ultra, while repetitive execution tasks are assigned to Lightning. The company claims its output speed can reach up to four times that of comparable models. On PinchBench, Lightning achieves an accuracy rate of 86%; under similar accuracy conditions, it completes 10,000 tasks 30% faster than Qwen3.6 35B. The model employs an MoE architecture and is specifically trained for Agent Harness, incorporating multi-token prediction and speculative decoding. NVIDIA provides weights in BF16 and NVFP4 formats, enabling deployment on local devices such as RTX 5090 and DGX Spark, and supports llama.cpp, Ollama, LM Studio, and Unsloth. It also emphasizes customizability: NVIDIA has opened access to the weights, partial training data, and training recipes, allowing enterprises to perform further fine-tuning for specialized tasks such as coding, security, and legal applications. In official demonstrations, Lightning showed significant performance gains across multiple professional tasks after fine-tuning.