NVIDIA Nemotron 3.5 Lightning: A fast 30B MoE for agent execution

NVIDIA Nemotron 3.5 Lightning is an open-source 30B mixture-of-experts (MoE) model that activates just 3B parameters per token, built specifically to power the high-volume execution layer of always-on AI agents.

Running autonomous agents eats up an enormous amount of tokens. Most of those tokens aren't spent pondering deep philosophical questions; they are wasted on routine chores like tool calls, git pulls, format checks, and subagent handoffs. Firing up a frontier model for every minor action balloons latency and drains developer budgets.

NVIDIA's solution splits the labor: let frontier reasoning giants handle the high-level roadmap, and let Nemotron 3.5 Lightning do the heavy lifting in the trenches. Alongside the model, NVIDIA also dropped NeMo Switchyard, an open-source routing library that directs incoming tasks to the best-suited model automatically.

What it does and how it runs so fast

Nemotron 3.5 Lightning specializes in fast, accurate task execution. It packs a 1M token context window and comes fine-tuned for popular agent harnesses such as OpenClaw and Hermes Agent.

Speed is the main selling point here. Beyond routing tokens to only 3B active parameters at a time, the model features built-in speculative decoding. NVIDIA trained it with native multi-token prediction (MTP) and shipped two dedicated draft models: DSpark for DGX Spark and low-concurrency workloads, and DFlash for higher concurrency. It also ships with an NVFP4 quantized checkpoint right beside BF16.

On the Artificial Analysis Intelligence Index, Nemotron 3.5 Lightning sits squarely on the accuracy-versus-speed Pareto frontier, delivering up to 4x the output speed of similar-sized options. In PinchBench evaluations, it hit 86% accuracy while finishing 10,000 tasks 30% faster than Qwen3.6 35B at comparable performance.

Early hands-on reviews note that while it does not match top-tier dense models like Qwen3.6 in raw intelligence, its raw token throughput easily outpaces speedsters like Gemini 3.6 Flash on routine execution.

Hardware specs and open availability

Thanks to native NVFP4 quantization, the full model checkpoint shrinks to around 22 GB. That means you can run it locally on upper-mid to high-end hardware, including a GeForce RTX 5090 desktop, an NVIDIA DGX Spark box, or a Jetson device.

It is supported out of the box across community tools like Ollama, llama.cpp, LM Studio, and Unsloth. On data center stacks, it plugs straight into vLLM, SGLang, and TensorRT-LLM.

NVIDIA released the model weights, training recipes, and data under the permissive OpenMDW-1.1 license. Developers can fine-tune it via NeMo Automodel or run reinforcement learning pipelines with NeMo Gym. The release also includes Nemotron-RL Agentic Terminal Pivot, an open dataset focused on terminal and coding agent skills.

Weights are available now on Hugging Face and ModelScope, and interactive endpoints are live on build.nvidia.com and OpenRouter.

Why it matters

If you build continuous agent workflows or want an affordable on-device worker to monitor logs and validate code changes, Nemotron 3.5 Lightning is an easy pick. Do not expect it to replace your primary reasoning engine for complex synthesis, but for running thousands of mundane terminal commands and tool calls, the latency drop is massive.

This release shows NVIDIA pushing hard into model orchestration. By pairing a lightweight MoE executor with the Switchyard router, developers can route high-level plans to top-tier proprietary models while dumping routine execution down to Lightning. NVIDIA's internal tests show that this split cuts task completion costs to nearly one-third of using a frontier model alone, creating a very practical blueprint for cost-effective agent deployments.