NVIDIA Nemotron 3.5 Lightning

Initial verdict: it is worth piloting for teams that need controllable or local inference, but the available evidence is not strong enough to make it a default production model. Its stated focus is clear—speed, tool use, and long-running agents—rather than simply maximizing general chat quality. A small, task-specific evaluation is therefore more appropriate than assuming that its model size or launch attention guarantees better results for every workflow.

In practical terms, this is a 30B-parameter hybrid MoE language model with about 3B active parameters. The design is intended to reduce the amount of computation used per token while supporting tool-driven automation and persistent agent loops. NVIDIA’s social announcement says that the weights, data, and recipes are open and customizable. Posts from OpenRouter and Ollama indicate availability through their hosted or local ecosystems. NVIDIA Robotics also stated an average of 115 tokens/s on Jetson AGX Thor; that is an official social-media claim for a stated setup, not a stable throughput guarantee across hardware, quantization choices, prompts, or workloads.