NVIDIA has released Nemotron 3.5 Lightning, an open 30B mixture of experts model designed to run high volume tasks for autonomous software agents. Operating with just 3B active parameters, the new release aims to execute routing, tool calls, and code validation at a fraction of the cost of frontier reasoning engines. The model is available now for local deployment on hardware ranging from the GeForce RTX 5090 to enterprise datacenters.
Running always on software agents can quickly become expensive. If an agent sends every single code check or formatting request to a massive frontier model, token budgets vanish. Most of an agent's life is spent on routine tasks. This high volume execution layer is where the new model is designed to work.
Developers can use the new NeMo Switchyard library to split the labor. Complex planning routes up to frontier models like Nemotron 3 Ultra. Simple tool execution routes down to the lightning model. This division keeps computing costs low.
The speed of the model comes down to its mixture of experts architecture. While the system has a 30B total parameter capacity, an internal router directs each token to only a few chosen experts. This process runs only 3B active parameters per token. It gives you the accuracy of a massive model with the speed of a small one.
The performance gains are clear in standardized benchmarks. In PinchBench testing, the model reached 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar precision levels. It completes actual work faster, not just generating empty tokens.
NVIDIA achieved this speed without compromising output accuracy. The model underwent pretraining to bake multi token prediction directly into the architecture. It also ships with specialized DSpark and DFlash draft models to help optimize serving speeds depending on your concurrency needs. For storage efficiency, the model includes an NVFP4 checkpoint that runs smoothly on desktop setups and enterprise GPUs alike.
Local deployment is highly accessible. The model is small enough to run on local systems including NVIDIA Jetson, the GeForce RTX 5090, and DGX Spark. It also works with standard developer tools like Ollama, llama.cpp, Unsloth, and LM Studio. The weights are open and customizable, allowing developers to fine tune the model quickly on modest hardware.
