Nvidia Debuts Nemotron 3.5 Lightning for Fast AI Agents

Nvidia has introduced Nemotron 3.5 Lightning, a 30-billion parameter open Mixture-of-Experts (MoE) model featuring 3 billion active parameters. Built specifically to handle the high-volume execution layer of autonomous AI agents, the model targets routine tasks like tool calls, result validation, and subagent delegation. By utilizing an MoE architecture, the model routes tokens to specific experts, delivering the capabilities of a larger dense model at a fraction of the compute cost. It is released under the permissive OpenMDW-1.1 license, providing weights, training data, and recipes for customization.

In performance evaluations, Nemotron 3.5 Lightning achieved up to four times the output speed of similarly sized models. On the Artificial Analysis Intelligence Index, which aggregates nine benchmarks across coding, agentic tasks, and scientific reasoning, the model established a new position on the accuracy-speed Pareto frontier. Furthermore, on the PinchBench benchmark, the model achieved 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at a comparable accuracy level.

To achieve these speeds without sacrificing accuracy, Nvidia integrated speculative decoding with multi-token prediction (MTP) during pretraining. The release includes two draft models: DSpark, optimized for DGX Spark and low-concurrency workloads, and DFlash. The model also ships with BF16 and NVFP4 checkpoints, utilizing specialized kernels that run on Nvidia Blackwell, Hopper, and Ampere GPUs. This allows practitioners to deploy the model locally on hardware like the GeForce RTX 5090 and Jetson, or scale up to data centers.

For developers building multi-model systems, Nvidia introduced NeMo Switchyard, a routing library that directs tasks to the most efficient model. This allows complex planning to go to frontier models like Nemotron 3 Ultra, while routing high-volume execution tasks down to Nemotron 3.5 Lightning. The release also includes the Nemotron-RL Agentic Terminal Pivot dataset to assist with training coding agents, and supports popular agent frameworks like OpenClaw, Hermes Agent, and the NemoClaw security stack.