NVIDIA ships Nemotron 3.5 Lightning as the execution engine for agent fleets
The open model handles the high-volume grunt work of long-running agents while a new open-source router sends the hard thinking to frontier models. NVIDIA claims the split slashes task costs.
NVIDIA released Nemotron 3.5 Lightning on Monday, an open 30 billion parameter mixture-of-experts model with just 3 billion active parameters, built for the unglamorous execution layer of always-on AI agents. Alongside it the company shipped NeMo Switchyard, an open-source routing library that decides, per task, whether work goes to Lightning or up to a frontier model.
The design bet is that agent fleets spend most of their tokens on repetitive, specialized work: tool calls, data processing, validation, and inter-agent messaging. In the launch announcement, NVIDIA says the hybrid architecture, which mixes Mamba-2 layers, MoE blocks, and selective attention, delivers up to 4x the output speed of similar-sized open models. A companion technical post reports that in a 10,000-task benchmark run the model finished about 30 percent faster than comparable alternatives at similar accuracy. Those are NVIDIA's own measurements, but the shape of the claim matters: this is a model graded on task completion, not chat quality.
Routing is the actual product
Switchyard is the more consequential release for orchestration teams. The library routes requests across an organization's mix of open, proprietary, and NVIDIA models, sending complex planning to frontier models and high-volume execution to Lightning. NVIDIA's internal testing found the blended setup maintained frontier-level accuracy while cutting task completion cost to roughly a third of running a frontier model alone. That number is unverified outside NVIDIA, but it formalizes what sophisticated agent shops already do by hand, and it slots into the security and management stack the company has been assembling around NemoClaw.
Lightning was also trained against the harnesses agents actually run in, with tuning for OpenClaw and Hermes Agent among others, and it ships in BF16 and NVFP4 quantized checkpoints. The weights, training data, and recipes are released under the permissive OpenMDW-1.1 license, so operators can fine-tune the execution layer on their own task distributions with LoRA or full supervised runs through NeMo tooling.
The economics of the execution tier
For GaaS operators, the two-tier pattern is becoming the default architecture. Frontier models like Claude Opus 5 or GPT-5.6 Sol plan and supervise; a cheap, fast, open executor grinds through the volume. Early adopters are already specializing the model, with CrowdStrike, Harvey, CodeRabbit, and Lila Sciences customizing Lightning for security, legal, code review, and scientific work respectively, according to NVIDIA.
Distribution is wide from day one: Hugging Face, ModelScope, OpenRouter, and build.nvidia.com host the weights, a NIM microservice covers data center deployment, and the model runs locally on DGX Spark, Jetson, and GeForce RTX 5090 hardware through Ollama, llama.cpp, and vLLM. That spread is the point. NVIDIA does not need Lightning to win benchmarks; it needs always-on agents to become a permanent, hardware-hungry workload with a routing layer that treats every model, including the frontier ones, as interchangeable capacity. Buyers get cheaper fleets today. Model vendors get commoditized tomorrow.