NVIDIA has added Nemotron 3.5 Lightning to its Nemotron 3 family of open models[1]. It is a 30-billion-parameter mixture-of-experts (MoE) model aimed squarely at handling the small, repetitive jobs inside always-on agents at high speed. Alongside it, NVIDIA released NeMo Switchyard, an open source routing layer that sends each step of a workflow to the most suitable model.

Always-On Agents Run as Systems of Models

The premise behind this launch is that modern agents operate as ensembles rather than a single model[1]. A strong reasoning model such as Nemotron 3 Ultra or GPT-5.6 plans and orchestrates the workflow, while smaller, faster models take on the individual tasks: code review, tool use, security alert monitoring, and answering billing questions.

Nemotron 3.5 Lightning sits on the small-and-fast side of that split. It follows Nemotron 3 Nano and is presented as the highest-efficiency model in its class for long-running agentic workloads[1].

Up to 4x Faster Output, 30 Percent Faster Task Completion

NVIDIA cites up to 4x faster output speed and 30 percent faster agentic task completion compared with other models in its class[1][2]. For agents designed to run continuously, per-call latency translates directly into operating cost and perceived responsiveness, so tightening that number makes practical sense.

The model was developed with contributions from the Nemotron Coalition, whose members supplied evaluation methodologies, inference software, and datasets[1]. Because the weights are open, teams can post-train it with NVIDIA NeMo on their own domain data, tools, and workflows to raise accuracy on specific tasks.

Industry-specific customizations are already under way[1]. CrowdStrike is tuning it for cybersecurity, Harvey is working with Trajectory on legal services, and CodeRabbit is working with Baseten on code review. Lila Sciences is improving reasoning for agentic tasks across the physical and life sciences, while Fastino Labs reports leading accuracy for software development, finance, and healthcare workloads.

From a Desktop RTX PC to the Data Center

The breadth of supported hardware is another selling point. The model runs locally on NVIDIA RTX PCs, DGX Spark, DGX Station, and Jetson, and scales out to edge AI devices, RTX PRO workstations, data centers, and cloud environments[1]. Running it locally also means keeping data in-house.

For local deployment, NVIDIA worked with vLLM, Ollama, llama.cpp, and LM Studio, offering the model in both NVFP4 and GGUF formats[2]. Unsloth provides day-one support with optimized and quantized builds through Unsloth Studio. Supported hardware extends to OEM GB10 systems and GB300 deskside machines, with Blackwell systems available from Acer, ASUS, Dell Technologies, Exxact, GIGABYTE, HP, Lenovo, MSI, and Supermicro.

Transparency around training also gets a mention. With every Nemotron launch, NVIDIA publishes as much of the training data and technique as licensing permits, and this time it also released Nemotron-RL-Agentic-Terminal-Pivot, the agentic reinforcement learning dataset used to post-train the model for coding agent capabilities[1]. That supports traceability and auditing, and the dataset can be used to train other models.

NeMo Switchyard Automates the Question of Which Model to Call

NeMo Switchyard is an open source model routing library for AI agents[1]. Some models are better at coding, some at reasoning, some at lightweight work, and some are optimized to run locally. Defaulting to a single model means overspending or losing quality, while wiring up routing by hand becomes integration work that slows a deployment down.

Switchyard automates that decision for each step of a workflow. Developers can tune or swap the routing algorithm to match priorities such as quality, latency, and cost, and adopt it without rewriting their applications[1]. In NVIDIA internal benchmarks, Switchyard maintained frontier-level accuracy while cutting task completion cost to nearly one-third of Opus 4.8 alone[1].

Partner results are specific[1]. Boomi evaluated Switchyard across 5 routing capabilities, achieving 100 percent domain-routing accuracy, sending 59 percent of traffic to a 5x faster fine-tuned model, and reducing later-turn latency by 21 percent. LangChain cut cost by 74 percent across 145 multi-turn Deep Agents tasks by routing only 7 percent of calls to a frontier model, at a 6 percent accuracy tradeoff. Ramp matched a frontier model's performance on Ramp SWE-Bench while cutting cost by 58 percent and runtime by 33 percent.

Elsewhere, Cognition integrated the staged router into Devin Desktop and reduced mean cost by 28 percent on FrontierCode Main. Classmethod is running opencode and Fireworks workloads internally with an initial 27 percent cost reduction, and Cadence improved efficiency by 9.9 percent with the ChipStack AI Super Agent on a formal verification use case. Kong delivers routing natively through Kong AI Gateway, LiteLLM is adding it as a proxy-layer plug-in, Nous Research has integrated it into Hermes, and Siemens is benchmarking it in its Fuse EDA AI Agent.

Where to Get It

Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice, as well as through NVIDIA Cloud Partners and a range of post-training and inference platforms[1]. NeMo Switchyard is on GitHub, with partner platform integrations coming soon.

Summary

Nemotron 3.5 Lightning is not competing for the top of the intelligence charts. It is built to handle, cheaply and quickly, the hundreds of small calls an agent makes in the course of its work. Combine the 4x output speed with automatic routing through NeMo Switchyard, and teams currently sending everything to a frontier model have a credible path to much lower cost. Given that the weights are open and the model runs on a desktop RTX PC, the deciding factor will likely be whether a team can post-train it on its own data.

Source[1] https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/

Source[2] https://blogs.nvidia.com/blog/local-ai-open-source-models-agents-nemotron/