Nvidia Releases Nemotron 3.5 Lightning, a 30-Billion-Parameter Open Model Built for High-Volume AI Agent Execution

Overview

Nvidia has released Nemotron 3.5 Lightning, an open 30-billion-parameter mixture-of-experts model with 3 billion active parameters, purpose-built for what the company calls the “execution layer” of AI agents — the high volume of tool calls, result validation, and subagent delegation that long-running agents spend most of their time on, according to NVIDIA’s developer blog. The company released the model alongside NeMo Switchyard, an open-source routing library designed to send individual agent tasks to whichever model — Lightning or a larger frontier system — best fits the job, according to NVIDIA’s developer blog.

What We Know

Nemotron 3.5 Lightning is the smallest member of the Nemotron 3 model family, which previously consisted of three sizes — Nano, Super, and Ultra — launched in December, according to SiliconANGLE. Nvidia positions the new model as the counterpart to larger reasoning models such as Nemotron 3 Ultra: frontier models handle orchestration and complex planning, while Lightning handles the high-volume execution work underneath, according to NVIDIA.

On speed, Nvidia and SiliconANGLE both describe the model as delivering up to four times the output speed and 30% faster agentic task completion compared with other models in its weight class, according to NVIDIA and SiliconANGLE. In an internal benchmark, Nvidia says the model reaches 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy, according to NVIDIA.

The model was trained with speculative decoding using multi-token prediction — a technique also used in Nemotron 3 Super and Nemotron 3 Ultra — and ships with two draft models, DFlash and DSpark, for further inference optimization, according to NVIDIA. Nvidia also describes the model as “harness-optimized,” trained for popular agent harnesses including OpenClaw and Hermes Agent so that agents make more accurate calls at lower latency on high-volume tasks, according to NVIDIA. Nvidia says the model deploys anywhere from a DGX Spark to a data center, and is released under a license Nvidia calls “as permissive as possible,” OpenMDW-1.1, with weights available on Hugging Face and ModelScope and hosted access through build.nvidia.com and OpenRouter, according to NVIDIA. Deployment guides cover vLLM, SGLang, and TensorRT-LLM, according to NVIDIA.

Nvidia says one early adopter, CodeRabbit, trained a router agent for $85 in around two hours using a single training epoch, an example the company points to as evidence of how quickly partners can customize the model, according to SiliconANGLE. SiliconANGLE reports the customization example was highlighted by Kari Briski, Nvidia’s vice president of generative AI, according to SiliconANGLE.

Alongside Lightning, Nvidia introduced NeMo Switchyard, an open-source model-routing library built around a provider-agnostic SDK called switchyard-libsy, which represents requests, defines the models available to a system, and manages calls to whichever model is selected, according to NVIDIA. At runtime, a router evaluates each request and its available context, then sends the work to the model that best suits the task’s requirements, constraints, and policies, according to NVIDIA. The library offers both “tuning-free” routers — an LLM classifier, a stage router, and an escalation router — and a “tunable” prefill router, according to NVIDIA.

Nvidia published two benchmark results for the routing approach. Across five runs, routing requests between Nemotron 3.5 Lightning and Claude Opus 4.8 with the escalation router delivered a 74% cost reduction compared with a frontier-only baseline, sending just 7% of calls to the frontier model at a measured roughly 6-point accuracy tradeoff, according to NVIDIA. On Cognition’s FrontierCode Main benchmark for production-grade coding tasks, an implementation routing between Opus 5 and Kimi K2.7 achieved 50.6% accuracy at a $3.11 mean cost — within 2.8 percentage points of Opus 5’s accuracy at roughly 28% lower mean cost, according to NVIDIA. Nvidia lists Kong, Classmethod, Boomi, Cadence, Siemens, Ramp, and Nous Research as ecosystem partners working on intelligent routing integration, according to NVIDIA, while SiliconANGLE additionally names Cognition AI as a partner collaborating with Nvidia on the effort, according to SiliconANGLE.

What We Don’t Know

Nvidia’s blog post does not disclose a context window length for Nemotron 3.5 Lightning. It is also unclear how the routing cost and accuracy figures for Nemotron 3.5 Lightning paired with Claude Opus 4.8, or for Opus 5 paired with Kimi K2.7, would generalize outside the specific LangChain and Cognition benchmark setups Nvidia cites.