Nvidia launches Nemotron 3.5 Lightning for faster AI agents
Nvidia's Nemotron 3.5 Lightning aims to make AI agents faster with efficient execution, lower latency and flexible deployment across devices.
Nvidia has introduced Nemotron 3.5 Lightning, an open AI model designed to help long-running AI agents complete routine tasks faster and with lower latency.
Announced on August 11, 2026, the model targets the execution layer of AI agent workflows, where systems repeatedly call tools, check outputs, format information, and delegate tasks.
Instead of relying on one large model for every task, developers are combining different models, using powerful reasoning systems for complex decisions and smaller models for repetitive work.
A smaller model built for repeated tasks
Nemotron 3.5 Lightning is a 30B parameter Mixture-of-Experts model with 3B active parameters. A Mixture-of-Experts model activates only selected sections for each task, allowing it to use fewer computing resources while retaining the capacity of a much larger model.
Nvidia says Lightning is designed for high-volume, low-latency workloads in always-on AI agents. It has also been trained for agent frameworks including OpenClaw and Hermes Agent, with a focus on accurate tool calls and faster execution.
The model uses techniques including speculative decoding, where a system generates possible tokens in advance and then verifies them more efficiently. Nvidia is also providing draft models called DSpark and DFlash for different inference requirements.
Speed becomes a key part of agentic AI
Nvidia says Nemotron 3.5 Lightning can deliver up to four times the output speed of comparable models. The company says its model reached 86% accuracy on PinchBench, completing 10,000 tasks 30% faster than Qwen3.6 35B at a comparable level of accuracy.
The model is available in BF16 and NVFP4 formats. NVFP4 is a lower-precision format designed to reduce the computing and memory requirements of AI inference on supported Nvidia hardware.
Nvidia states Lightning can be deployed across different environments, including data centres, DGX Spark and local systems such as Jetson and GeForce RTX 5090.
Routing each task to the right model
Nvidia has also introduced NeMo Switchyard, a library designed to route tasks between different AI models.
The idea is simple: a complex planning task can be sent to a larger reasoning model, while repetitive execution can be handled by Nemotron 3.5 Lightning.
Developers can therefore build agent systems without forcing one model to handle every part of a workflow.
For businesses, this approach could reduce computing costs and improve response times while allowing AI agents to operate across cloud and local infrastructure.
As AI agents move towards completing longer and more complex workflows, faster execution could become just as important as stronger reasoning. Nvidia's latest model reflects that shift, focusing on the less visible but essential work that keeps autonomous AI systems running.