As artificial intelligence transitions from chatbot interfaces toward autonomous agents, open-source models are increasingly meeting enterprise demands for control over deployment location, operational parameters, and model evolution.

NVIDIA has announced the expansion of its Nemotron 3 model family with the release of Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model positioned as the highest-efficiency option in its class for long-running agentic AI workloads. The launch follows Nemotron 3 Nano and continues NVIDIA's effort to refine open models for improved accuracy and speed.

Alongside the new model, NVIDIA is introducing NeMo Switchyard, an open-source library designed for intelligent routing within popular agent development tools. The library enables enterprises to construct routers tailored to their requirements, automatically directing each request to the most capable and appropriate model for the task without requiring application rewrites.

Together, Nemotron 3.5 Lightning and NeMo Switchyard are intended to give organizations greater control over AI deployment, runtime location, and operational efficiency across PCs, workstations, data centers, and cloud infrastructure as agentic systems move beyond single-model chat experiences and into workflow-level automation.

Always-On Agents and the System-of-Models Architecture

Modern agentic systems — characterized as always-on agents — increasingly function as ensembles of specialized models, each optimized for distinct tasks. NVIDIA designed its Nemotron open models for this architecture. A frontier reasoning model such as Nemotron 3 Ultra or GPT-5.6 can plan and orchestrate an overall workflow, while smaller specialized models like Nemotron 3.5 Lightning execute targeted operations such as code review, tool usage, security alert monitoring, and billing inquiries.

Nemotron 3.5 Lightning: Built for High-Volume Specialized Tasks

Nemotron 3.5 Lightning is a fully customizable open model developed for the high-volume tasks that power always-on agents. The model was built with contributions from the Nemotron Coalition, whose members supplied evaluation methodologies, inference software, and datasets to advance the model's capabilities.

According to NVIDIA, Nemotron 3.5 Lightning delivers up to 4x faster output speed, resulting in 30% faster agentic task completion compared with other models in its class. Because the model is open and customizable, organizations can post-train it using NVIDIA NeMo on their own domain data, tools, and workflows to enhance accuracy for specialized use cases.

Several AI leaders across industries are already customizing Nemotron 3.5 Lightning for their specific workloads:

• CrowdStrike is applying the model to cybersecurity use cases.
• Harvey, working with Trajectory, is deploying it for legal services.
• CodeRabbit, in collaboration with Baseten, is using it for code review.
• Lila Sciences is leveraging the model to improve reasoning for agentic tasks across physical and life sciences.
• Fastino Labs has customized the model and reports leading accuracies for software development, finance, and healthcare workloads.

Nemotron 3.5 Lightning also provides organizations with privacy and deployment flexibility. It can run on local AI systems — including NVIDIA RTX PCs, NVIDIA DGX Spark, NVIDIA DGX Station, and NVIDIA Jetson — helping users maximize existing infrastructure. It can also scale across edge AI devices, NVIDIA RTX PRO workstations, data centers, and cloud environments for enterprise applications. For high-volume, specialized tasks that demand fast responses, the model can run locally or on-premises, which may matter for organizations balancing latency, privacy, and infrastructure reuse.

Consistent with prior Nemotron releases, NVIDIA is publishing as much of the training data and methodology as licensing permits, supporting traceability, auditing, and further model training. Alongside Lightning, NVIDIA is also releasing Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement learning dataset used to post-train the model for coding agent capabilities.

NeMo Switchyard: Efficient Model Routing for AI Agents

Different models excel at different tasks — some are optimized for coding, others for reasoning, lightweight operations, or local execution for privacy. Relying on a single default model can lead to either excessive spending or quality degradation, while manual routing introduces integration overhead that can slow deployments.

NeMo Switchyard addresses this by automatically routing prompts to the most capable and efficient model for each step of an agent workflow. Developers can tune the router with different algorithms to align with priorities such as quality, latency, and cost. NVIDIA's internal benchmarks indicate that NeMo Switchyard maintains frontier-level accuracy while reducing task completion cost to approximately one-third of using Opus 4.8 alone.

For teams building production agents, that routing layer is a practical middle ground between one-size-fits-all inference and custom orchestration for every application. NVIDIA is collaborating with partners across the AI ecosystem to integrate intelligent model routing into existing developer tools and platforms. Early results include:

• Boomi: Evaluated Switchyard across five routing capabilities, achieving 100% domain-routing accuracy, directing 59% of traffic to a 5x faster fine-tuned model, and reducing later-turn latency by 21%.
• Cadence: Improved efficiency by 9.9% using the ChipStack AI Super Agent for a formal verification use case.
• Classmethod: Running opencode and Fireworks workloads with NeMo Switchyard internally, with initial testing showing a 27% cost reduction while maintaining quality.
• Cognition: Integrated the NeMo Switchyard staged router into Devin Desktop for NVIDIA internal use, achieving near-frontier performance on FrontierCode Main while reducing mean cost by 28% compared to routing all requests to a single frontier model.
• Kong: Delivers routing with NeMo Switchyard natively through Kong AI Gateway.
• LangChain: Achieved 74% lower cost across 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model, with a 6% accuracy tradeoff.
• LiteLLM: Adding NeMo Switchyard as a plug-in to its proxy layer, enabling developers to access routing benefits without modifying their existing stack.
• Nous Research: Integrated NeMo Switchyard into Hermes to provide developers with an easy-to-configure routing system.
• Ramp: Used NeMo Switchyard to match a frontier model's performance while reducing costs by 58% and runtime by 33% in Ramp SWE-Bench.
• Siemens: Benchmarking the library to improve efficiency in its Fuse EDA AI Agent.

Availability

Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice, as well as through a broad ecosystem of NVIDIA Cloud Partners, post-training platforms, inference platforms, and cloud service providers. NeMo Switchyard is available on GitHub and is expected to arrive on partner platforms in the near future.