AI agents have a hidden cost problem. Once an agent finishes its "thinking," most of its runtime is spent on unglamorous work: calling tools, validating results, formatting outputs, and delegating to subagents. Sending every one of those routine steps to a massive frontier model burns money and adds latency. On August 11, 2026, Nvidia released its answer: Nemotron 3.5 Lightning, a free, open AI model built specifically for the high-volume execution layer of always-on agents.

What Nemotron 3.5 Lightning Actually Is

Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts (MoE) model, but only about 3 billion parameters are active for any given request. That design gives it some of the capacity of a much larger system while keeping compute demands closer to those of a small model. Under the hood, it uses a hybrid architecture that combines Mamba-2 layers with MoE blocks and attention, and it supports a context window of up to one million tokens, which matters for agents that accumulate long histories of tool calls and intermediate results.

The model was distilled from Nemotron 3 Ultra, Nvidia's 550-billion-parameter frontier open model, and it expands the Nemotron 3 family the company introduced in December, which originally shipped in Nano, Super, and Ultra sizes. Lightning is the smallest and most efficient member of that family, tuned for popular agent harnesses and for specialized tasks such as code review and tool use, plus jobs like monitoring security alerts inside larger multi-agent systems.

The Numbers Behind the Launch

Nvidia's pitch is heavily quantitative. The company reports that Lightning generates output up to 4x faster than open models in its class and completes agentic tasks about 30% faster. In one company-published evaluation, the model recorded 86% accuracy on an agentic benchmark and finished a batch of 10,000 tasks 30% faster than a comparable 35-billion-parameter open model at similar accuracy. Nvidia also places Lightning on the accuracy-speed frontier for small open models on a widely followed independent intelligence index that aggregates nine evaluations spanning coding and scientific reasoning as well as agentic tasks.

Most of these figures come from the vendor, so independent verification will follow in the coming weeks.

Still, the early customization numbers stand out. One code-review company reportedly fine-tuned Lightning into a working router agent using a standard recipe, a single training epoch, roughly two hours of compute, and about $85 in cost.

Why This Matters for You

For developers and enterprise teams, the practical appeal is control and cost. The model is free to download, use, modify, and redistribute without asking Nvidia's permission, and the release goes further than most: the weights are published under a permissive open license, along with the training data and recipes. That means you can fine-tune it with LoRA or full supervised fine-tuning and run reinforcement learning on top of it. You can also reproduce or audit how it was built. An open agentic RL dataset used to train some of its coding-agent skills ships alongside it.

AI enthusiasts running models locally get a model designed to fit on a single GPU. Nvidia highlights that it runs on consumer and edge hardware, including the GeForce RTX 5090 and Jetson devices, plus compact desktop AI systems. You can power a personal always-on agent on your own machine and train it to write in your preferred style, with your data staying local.

CrowdStrike and CodeRabbit have already tested and customized the model for their own workloads, as has Harvey. It's available through the major open ecosystems developers already use, including popular local runtimes and standard inference frameworks, as well as Nvidia's own site and model hubs.

The Other Half of the Story: NeMo Switchyard

Lightning didn't ship alone. Nvidia simultaneously released NeMo Switchyard, an open-source routing library that automatically directs each step of an agent workflow to the most capable and cost-efficient model available, whether that's a small specialist like Lightning or a large frontier reasoner for complex planning.

The routing math is where the economics get interesting. Nvidia's internal benchmarks show Switchyard maintaining frontier-level accuracy while cutting task-completion costs to roughly one-third of what a leading proprietary frontier model costs on its own. Real-world pilots echo that: fintech company Ramp matched a frontier model's performance while reducing costs by 58% and runtime by 33%, and LangChain achieved 74% lower costs on multi-turn agent tasks by routing only 7% of calls to a frontier model, accepting a modest 6% accuracy tradeoff.

The Bigger Picture

The release reflects a clear industry shift: agentic AI is moving from single monolithic models toward ensembles of specialized ones, where a frontier model plans and orchestrates while smaller, cheaper models execute. Nvidia is betting that the future of agents looks like a well-run team, not a lone genius, and it's giving developers the open building blocks to assemble that team on their own terms, on their own hardware.