Published:
Nvidia released Nemotron 3.5 Lightning and NeMo Switchyard on August 11, 2026, a 30-billion-parameter mixture-of-experts model and an open-source routing library built to run enterprise AI agents at a fraction of frontier-model prices. In LangChain testing, NeMo Switchyard cut task costs by 74% against running Anthropic’s Opus 4.8 alone, sending only 7% of calls to the frontier model and losing roughly 6 points of accuracy. Nemotron 3.5 Lightning completes 10,000 tasks 30% faster than Qwen3 35B at comparable accuracy and scores 86% on PinchBench. Nvidia is also publishing the post-training datasets and recipes it used to build the models.
What Nemotron 3.5 Lightning Is Built to Do
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model specialized for high-volume, long-running agent tasks rather than open-ended reasoning. The model keeps only 3 billion of its 30 billion parameters active at inference, which is why it runs on a single PC GPU and generates tokens up to 4 times faster than comparable open models.
Three architecture choices produce that throughput: speculative decoding with multi-token prediction, NVFP4 quantization, and two draft models named DSpark and DFlash. Nemotron 3.5 Lightning is a distinct model from Nemotron 3 Ultra, Nvidia’s December 2025 release, not a version bump — Ultra targets frontier-grade reasoning, Lightning targets throughput per dollar.
“What we’re hearing is that Lightning is remarkably easy to customize.” — Kari Briski, VP of Generative AI, Nvidia (SiliconAngle, August 11, 2026)
How NeMo Switchyard Produces the 74% Saving
NeMo Switchyard is an open-source library that routes each individual step of an agent workflow to the cheapest model capable of handling it, escalating to a frontier model only when the step demands it. In Nvidia’s LangChain test, that routing sent 7% of calls to Opus 4.8 and the remaining calls to cheaper models, cutting total task cost to nearly one-third of running Opus 4.8 alone.
The economics matter because agent workflows multiply cost by step count. An enterprise running an always-on agent pays frontier prices for every retrieval, every tool call, and every intermediate summary, most of which need no frontier reasoning. Routing turns that flat premium into a 7%-of-calls premium.
The accuracy trade is explicit: the same test lost approximately 6 accuracy points against the all-Opus baseline. Nvidia measured the 74% figure in one LangChain configuration, and the saving any given workload sees depends on how many of its steps genuinely require frontier reasoning. Teams evaluating the best AI agents for business tasks should benchmark their own step mix before assuming the headline number transfers.
“Depending on your routing strategy, it wants to choose the best model.” — Kari Briski, VP of Generative AI, Nvidia, on NeMo Switchyard (SiliconAngle, August 11, 2026)
What Adoption Costs and Who Has Signed On
CodeRabbit built a specialized router agent for $85 in about two hours, using Nvidia’s auto model recipe and training for a single epoch, according to Kari Briski. That figure sets the practical floor for customization: a small team can produce a working routing agent for the price of a monthly SaaS seat.
Seven named partners have adopted NeMo Switchyard, including Boomi, Cadence Design Systems, LangChain, Cognition AI, Kong, Nous Research, and Siemens. The LangChain integration is the strategically important one, because most enterprise AI teams already orchestrate agents through LangChain, and NeMo Switchyard routes across whatever model mix a company runs — open, proprietary, or Nvidia’s own. Nvidia has not confirmed the commercial licensing terms of the open-source release.
For Context: Our Nvidia and AI Agent Cost Coverage
Nemotron 3.5 Lightning extends a run of releases aimed at the cost of running agents in production.
- Nvidia’s previous Nemotron 3 Ultra — the December 2025 foundation model this release builds on.
- Echo AI matches Fable quality for one-third the cost — cost-efficient AI models pursuing the same price gap.
- Cloudflare Kitesurf cuts agent costs 3.7x — other tools cutting AI agent costs at the infrastructure layer.
Our Take: Nvidia is turning itself into enterprise AI middleware — the chip, the model, the router, and the training recipe. NeMo Switchyard is the piece to watch, because a routing layer that sits inside LangChain gets a vote in every agent workflow regardless of which frontier lab wins the quality race. The 74% number will not survive contact with every workload, but the 6-point accuracy trade is a decision most operations teams will take.
