NVIDIA has released Nemotron 3.5 Lightning, a specialized AI model designed to handle the repetitive, high-volume work that dominates long-running AI agents, delivering up to 30% faster task completion while maintaining accuracy comparable to larger models. The 30-billion parameter mixture-of-experts model with 3 billion active parameters represents a shift in how developers architect AI agent systems, splitting work between a large reasoning model for planning and a smaller, faster model for execution.

Why Do AI Agents Need a Dedicated Execution Model?

Long-running AI agents spend most of their time on repetitive tasks: making tool calls, validating results, formatting outputs, and delegating work to subagents. Using a large frontier reasoning model for every single execution step adds unnecessary cost and latency. Nemotron 3.5 Lightning is built specifically for this execution layer, allowing developers to reserve their most powerful models for complex planning and orchestration.

The model achieves its speed through a mixture-of-experts architecture, which works like a specialized team where each token is routed to only a few experts rather than activating the entire model. This means the model has the reasoning capacity of a much larger system while consuming far less computing power per token. The result is a model that can run on everything from a single desktop GPU to large data center clusters.

What Performance Gains Does Lightning Deliver?

On the PinchBench agentic task benchmark, Nemotron 3.5 Lightning reached 86% accuracy while completing 10,000 tasks 30% faster than Qwen 3.6, a competing 35-billion parameter model at similar accuracy levels. The model also ranks on the accuracy-speed Pareto frontier, a technical term meaning it achieves the best combination of speed and correctness compared to other open-source models in its size class.

NVIDIA achieved this speed without sacrificing accuracy through several inference optimization techniques. The model includes speculative decoding, a process where the model predicts multiple tokens at once and efficiently validates them, reducing the number of inference steps needed. NVIDIA also released two draft models, DSpark and DFlash, that work alongside Lightning to further accelerate inference depending on your workload.

How to Deploy Lightning in Your AI Agent Architecture

• Model Routing Strategy: Use NVIDIA NeMo Switchyard to intelligently route requests to the right model; send complex planning tasks to a frontier model like Nemotron 3 Ultra and execution tasks to Lightning, ensuring efficient token spending across your system.
• Fine-Tuning for Your Workload: Customize Lightning using LoRA or full supervised fine-tuning with NeMo Automodel, which trains faster and cheaper on modest hardware compared to larger models, allowing you to adapt the model to domain-specific tasks.
• Reinforcement Learning Integration: Train Lightning with NeMo RL and NeMo Gym using the open Nemotron-RL Agentic Terminal Pivot dataset, which includes real agentic reinforcement learning examples for coding and tool-use tasks.
• Quantization for Efficiency: Deploy the NVFP4 quantized checkpoint alongside the BF16 version to reduce memory requirements while maintaining accuracy across NVIDIA Blackwell, Hopper, and Ampere GPUs.

What Makes Lightning Different From Larger Models?

The core innovation is architectural: Nemotron 3.5 Lightning is purpose-built for agents rather than general-purpose reasoning. The model was trained specifically on popular agent harnesses like OpenClaw and Hermes Agent, meaning it learned to make more accurate tool calls and format results correctly for agent frameworks. This harness-optimized training is a key reason why Lightning outperforms larger models on agentic benchmarks despite having fewer parameters.

NVIDIA released the model weights, training data, and recipes under the permissive OpenMDW-1.1 license, allowing developers to inspect, modify, and fine-tune the model without restrictions. This open approach contrasts with proprietary models and gives developers full control over customization.

The emergence of Lightning reflects a broader shift in AI agent development: instead of using a single large model for everything, developers are building systems with multiple models optimized for different tasks. Frontier models like Nemotron 3 Ultra handle orchestration and complex reasoning, while Lightning handles the high-volume execution that dominates token budgets in always-on agents. This division of labor, enabled by tools like NeMo Switchyard, allows developers to build faster, cheaper agents without sacrificing reasoning quality.