Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model, and NeMo Switchyard, an open-source model-routing library, on August 11, 2026. Lightning delivers up to 4x faster output speed and 30% faster agentic task completion than peers in its size class, according to Nvidia's PinchBench results. Switchyard, meanwhile, cuts task cost to nearly one-third of running Anthropic's Opus 4.8 alone by directing each step of an agent workflow to the cheapest model that can handle it.
The two releases target the same shift: agent systems are moving from single-model calls to ensembles, where a frontier planner like Nemotron 3 Ultra or GPT-5.6 delegates subtasks to smaller specialists. Nemotron 3.5 Lightning is built for the specialist role — code review, tool use, security alerts, billing questions — where high throughput and low latency matter more than raw reasoning depth. The release follows Nemotron 3 Nano, which we covered last week when Nvidia shipped the 30B open MoE aimed at local deployment.
“As AI shifts from chatbots to autonomous agents, open models are serving market demands for full control over where AI runs and how it's deployed and evolves.”— Kari Briski, Nvidia, author of the announcement
Because Lightning is open and customizable, enterprises can post-train it on their own data using NVIDIA NeMo. CrowdStrike is fine-tuning it for cybersecurity workflows, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review. Lila Sciences is customizing it for physical and life-sciences reasoning; Fastino Labs is deploying variants across software development, finance, and healthcare.
Key facts
- 01Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model delivering up to 4x faster output and 30% faster agentic task completion versus peers.
- 02NeMo Switchyard, an open-source routing library, cut task completion cost to nearly one-third of running Opus 4.8 alone in Nvidia's internal benchmarks.
- 03LangChain hit 74% lower cost across 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model, accepting a 6% accuracy tradeoff.
- 04Ramp matched frontier-model performance on Ramp SWE-Bench while cutting cost 58% and runtime 33% using Switchyard.
- 05Nemotron 3.5 Lightning ships on Hugging Face, ModelScope, OpenRouter and build.nvidia.com; NeMo Switchyard is on GitHub.
Deployment flexibility is the other pitch. Lightning runs locally on Nvidia RTX PCs, DGX Spark, DGX Station, and Jetson, and scales up to RTX PRO workstations, data centers, and cloud. Nvidia is also publishing the training data and techniques where licensing allows, and shipping Nemotron-RL-Agentic-Terminal-Pivot — the agentic reinforcement-learning dataset used to post-train Lightning for coding agents.
NeMo Switchyard tackles the tokenomics problem underneath all of this. If an agent defaults to a frontier model for every call, cost balloons. If developers hand-code routing, integration slows the deployment. Switchyard automates the choice, and lets developers tune the routing algorithm to prioritize quality, latency, or cost.
“The technology routes prompts to the most capable and efficient model for each step of an agent workflow automatically, based on specific needs.”— Kari Briski, Nvidia, author of the announcement
The partner benchmarks Nvidia published are the most concrete data in the release. LangChain hit 74% lower cost across 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model, at a 6% accuracy tradeoff. Ramp matched frontier-model performance on Ramp SWE-Bench while cutting cost 58% and runtime 33%. Cognition integrated Switchyard's staged router into Devin Desktop for Nvidia's internal use, achieving near-frontier performance on FrontierCode Main while reducing mean cost 28% versus routing every request to a single frontier model.
Boomi's evaluation across five routing capabilities hit 100% domain-routing accuracy, sent 59% of traffic to a 5x faster fine-tuned model, and cut later-turn latency 21%. Cadence improved efficiency 9.9% using the ChipStack AI Super Agent for formal verification. Classmethod is running opencode and Fireworks workloads on Switchyard with 27% cost reduction in initial testing. Kong is exposing routing natively through its AI Gateway; LiteLLM is adding Switchyard as a plug-in to its proxy layer; Nous Research integrated it into Hermes; Siemens is benchmarking it inside its Fuse EDA AI Agent.
The commercial logic is straightforward: Nvidia sells more GPUs when more workloads run on Nvidia hardware, and both a smaller open model and a router that keeps traffic on cheaper endpoints expand the surface area for that. But the strategic move is more interesting. By publishing an open router that treats frontier models as one option among many — and by benchmarking against Opus 4.8 explicitly — Nvidia is nudging enterprises toward architectures where the frontier model is a fallback, not a default.
The caveats are real. A 6% accuracy hit is trivial in some workflows and disqualifying in others; the LangChain result assumes buyers can tolerate that curve. Nvidia's own benchmarks favor Nvidia's own outcomes, and PinchBench is not yet an industry-standard yardstick. And router quality depends heavily on how well the customer tunes it — a badly configured Switchyard could route the wrong prompts to the wrong models and erase the savings.
For the AI market, the release sharpens a trend the frontier labs would rather not accelerate. If routing libraries can extract most of the accuracy of Opus 4.8 or GPT-5.6 while sending the bulk of tokens to open models running on customer-owned hardware, the economics of paying frontier per-token rates for every call get harder to defend. Nvidia is betting the next phase of agent deployment looks less like one big model and more like a fleet — and it wants to sell both the fleet and the dispatcher.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
