Published: 12 August 2026
NVIDIA released Nemotron 3.5 Lightning on 11 August 2026, an open 30 billion parameter mixture of experts model with 3 billion active parameters built for the high volume execution layer of long running AI agents. Alongside it NVIDIA released NeMo Switchyard, an open source library that routes each step of an agent workflow to the best fit model. NVIDIA reports 86 percent accuracy on PinchBench while completing 10,000 tasks 30 percent faster than Qwen3.6 35B at comparable accuracy, and publishes the weights, training data and recipes under the OpenMDW-1.1 licence.
What is Nemotron 3.5 Lightning built for?
Long running agents spend most of their time on repetitive execution work rather than planning: tool calls, validating results and delegating to subagents. Running a frontier reasoning model for every one of those steps adds cost and latency. Nemotron 3.5 Lightning is built specifically for that execution layer. It is the smallest member of the Nemotron 3 family and follows Nemotron 3 Nano, and NVIDIA positions it to sit under a frontier planner such as Nemotron 3 Ultra or GPT-5.6 in a system of models, handling the routine calls that dominate an agent’s token budget.
The mixture of experts design is what makes that economical. A router sends each token to a small subset of the model’s experts, so only 3 billion of the 30 billion parameters run per token. That gives the capacity of a larger dense model at roughly the compute cost of a small one. NVIDIA also trained the model against popular agent harnesses, including OpenClaw and Hermes Agent, both supported by the NVIDIA NemoClaw open source security and management stack for always on agents.
Enterprises are already customising it for narrow domains. NVIDIA names CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, CodeRabbit with Baseten for code review, Lila Sciences for reasoning across physical and life sciences, and Fastino Labs for software development, finance and healthcare workloads. Because the model is small, fine tuning is cheaper and runs on more modest hardware than a frontier model would need.
Nemotron 3.5 Lightning benchmarks and technical specs
All benchmark figures come from NVIDIA’s own evaluations.
- On PinchBench, Nemotron 3.5 Lightning reaches 86 percent accuracy while completing 10,000 tasks 30 percent faster than Qwen3.6 35B at similar accuracy.
- NVIDIA also places the model on the accuracy versus speed Pareto frontier of the Artificial Analysis Intelligence Index, an index that combines nine evaluations covering agentic tasks, coding, scientific reasoning and general intelligence, and claims up to 4 times the output speed of similar sized models. NVIDIA has not published the full evaluation harness or prompt sets behind the PinchBench comparison.
The speed comes from two techniques. Multi token prediction was baked into the model during a dedicated pretraining stage, as with Nemotron 3 Super and Nemotron 3 Ultra, followed by a further phase that improved its accuracy. On top of that, NVIDIA ships two draft models for speculative decoding: DSpark, recommended for DGX Spark inference and low concurrency data centre workloads, and DFlash. Multi token prediction is the better fit at medium to high concurrency, with the optimal draft length shrinking as concurrency rises.
For precision, Nemotron 3.5 Lightning ships an NVFP4 checkpoint alongside a BF16 checkpoint, using the same NVFP4 kernels that power Nemotron 3 Ultra across NVIDIA Blackwell, Hopper and Ampere GPUs. The same file works in a data centre and on a desktop DGX Spark. NVIDIA is also publishing Nemotron-RL-Agentic-Terminal-Pivot, the agentic reinforcement learning dataset used to post train some of the model’s coding agent capabilities.
What NeMo Switchyard does and what it saves
NeMo Switchyard is the second half of the release: an open source model routing library for AI agents. It sends each prompt in an agent workflow to the most capable and efficient model for that step, across a developer’s own mix of open, proprietary and NVIDIA models, without requiring the application to be rewritten. Developers can tune or swap the routing algorithm to match their priorities on quality, latency or cost. In practice, plans route up to a frontier model and execution routes down to a model like Nemotron 3.5 Lightning.
NVIDIA’s internal benchmarks show Switchyard maintaining frontier level accuracy while cutting task completion cost to nearly one third of using Claude Opus 4.8 alone. Partner results are more specific. LangChain reports 74 percent lower cost across 145 multi turn Deep Agents tasks by routing only 7 percent of calls to a frontier model, at a 6 percentage point accuracy tradeoff. Ramp matched a frontier model’s performance on Ramp SWE-Bench while cutting cost by 58 percent and runtime by 33 percent. Cognition integrated the staged router into Devin Desktop for NVIDIA internal use and reduced mean cost by 28 percent on FrontierCode Main relative to routing everything to one frontier model.
Other partners report narrower gains. Boomi measured 100 percent domain routing accuracy across five routing capabilities, sent 59 percent of traffic to a fine tuned model 5 times faster than the default and cut later turn latency by 21 percent. Classmethod is running opencode and Fireworks workloads on Switchyard internally with early testing showing a 27 percent cost reduction at the same quality. Cadence improved efficiency by 9.9 percent on a formal verification use case. Kong delivers Switchyard routing natively through Kong AI Gateway, LiteLLM is adding it as a proxy layer plug in, and Nous Research has integrated it into Hermes.
Nemotron 3.5 Lightning availability, licensing and hardware
Nemotron 3.5 Lightning is available now on Hugging Face as nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, on ModelScope, through OpenRouter on a free tier, and on build.nvidia.com as an NVIDIA NIM microservice. NeMo Switchyard is on GitHub under NVIDIA-NeMo/Switchyard and is coming to partner platforms. Weights, training data and recipes are released under OpenMDW-1.1, which permits free download, adaptation and commercial modification with no licensing fee.
The model is small enough to run locally. NVIDIA lists NVIDIA Jetson, the GeForce RTX 5090, DGX Spark and DGX Station as supported local systems, scaling up through RTX PRO workstations to data centres and cloud. It runs through the standard local tooling as well, including LM Studio, llama.cpp, Ollama and Unsloth. NVIDIA worked with EXO Labs to characterise DGX Spark performance and places the model on the Pareto frontier for small open models on the EXO Labs local.ai leaderboard.
For customisation, NVIDIA points to LoRA or full supervised fine tuning with NeMo Automodel and NeMo Megatron Bridge, and reinforcement learning with environment based evaluations through NeMo RL and NeMo Gym. Hosted inference is available from Baseten, CoreWeave, Crusoe, DeepInfra, Fireworks AI, FriendliAI, GMI Cloud, Modal, Nebius and Together AI, among others. The model was developed with contributions from the Nemotron Coalition, whose members supplied evaluation methodologies, inference software and datasets.
For the full details you can visit the official NVIDIA announcement on Nemotron 3.5 Lightning and NeMo Switchyard.
