Nvidia on Tuesday launched two complementary products aimed at the cost-performance tradeoff that enterprises face when running AI agents at scale: Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model, and NeMo Switchyard, an open-source routing library that directs each step of an agent workflow to whichever model fits it best1,2.
The pairing targets a specific pain point. Sending every agent task to a frontier model drives up costs; building custom routing logic to offload simpler tasks to cheaper models creates its own engineering burden. Switchyard, written in Rust, lets developers define a pool of models and set routing criteria tuned for quality, latency, and cost, with APIs rather than bespoke plumbing.
Nvidia's headline performance claims center on speed and cost. The company says Lightning delivers up to 4x faster output than comparable models in its class and completes agentic tasks roughly 30% faster than Qwen3.6-35B at matching accuracy, according to PinchBench data supplied by Nvidia. On the same benchmark data, Lightning beats Gemma 4 26B's accuracy at a similar completion time. However, both Nvidia Nemotron models trail the similarly sized Gemma 4 31B from Google on the Artificial Analysis Intelligence Index, where Lightning scores 24.
When Lightning is paired with Switchyard and a mix of open models plus Anthropic's Opus 4.8, Nvidia says the system maintained frontier-level accuracy while cutting task-completion costs to roughly a third of running Opus 4.8 alone.
Nvidia shared results from nine companies testing Switchyard. LangChain reported a 74% cost reduction across 145 multi-turn Deep Agents tasks by routing just 7% of calls to a frontier model, with a 6% accuracy tradeoff. Ramp said it matched a frontier model's performance on its internal SWE-Bench while cutting costs 58% and runtime 33%. Cognition integrated Switchyard's staged router into Devin Desktop for internal use and reported near-frontier performance on FrontierCode Main while cutting mean cost 28% relative to routing everything to a single frontier model.
Nvidia's partner ecosystem for Switchyard splits into two tiers: agent frameworks that call Switchyard directly — Cognition, LangChain, and Nous Research — and LLM gateways with native Switchyard support, including Kong, LiteLLM, and OpenRouter. Kong ships Switchyard natively inside Kong AI Gateway.
Lightning extends the hybrid Mamba-Transformer, latent mixture-of-experts architecture Nvidia introduced with the Nemotron 3 family in December 2025. It was developed with contributions from the Nemotron coalition and has reasoning capabilities that come close to the much larger Nemotron 3 Super model. Nvidia gave customers early access to customize the model for specialized workflows, working with partners including CrowdStrike and CodeRabbit. Nvidia Senior Director Joey Conway said the company believes systems of models are the future of AI. Nvidia's Kari Briski noted during a press briefing that the ability to modify and optimize the model for specific workflows is what Nvidia hopes will set it apart.
Nvidia said its standard NeMo Auto model recipe, trained for one epoch, built into a working router agent for $85 in about two hours. The company is also releasing a full dataset for agentic reinforcement learning to train coding agents.
Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an Nvidia NIM microservice. NeMo Switchyard is available on GitHub.
ANALYSIS The launch positions Nvidia not just as a chip supplier but as an infrastructure layer for multi-model agent orchestration, competing directly with the routing and gateway tooling that startups and cloud providers have been building independently. The cost-reduction figures from early adopters like LangChain and Ramp, if they hold at production scale, could make Switchyard a default integration point for enterprises already running on Nvidia hardware.