NVIDIA's latest release is really an argument about architecture. Rather than push out a single larger model, the company shipped two pieces that only make sense together: Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model tuned for long-running agent work, and NeMo Switchyard, an open source router that decides which model in a stack should field any given request.
The reasoning behind the pairing is straightforward enough. Agentic systems increasingly run as ensembles, with a heavyweight planner such as Nemotron 3 Ultra or GPT-5.6 setting direction while cheaper specialists grind through code review, tool invocation, alert triage and routine queries, and NVIDIA is now selling into both halves of that shape.
Lightning: Small, Fast, Meant to Be Modified
NVIDIA claims up to 4x higher output speed for Lightning against class peers, which it translates into roughly 30 percent quicker completion of agentic tasks, citing PinchBench as supporting evidence. Speed is the pitch, but customization is the strategy: the weights are open, so teams can post-train on proprietary data using NeMo, and NVIDIA is publishing Nemotron-RL-Agentic-Terminal-Pivot, the reinforcement learning dataset behind the model's coding agent behaviour, alongside as much training methodology as licensing permits.
That openness has already produced a customer list worth reading as validation — CrowdStrike in security, Harvey with Trajectory in legal services, CodeRabbit with Baseten for code review, Lila Sciences across physical and life sciences, and Fastino Labs spanning software, finance and healthcare. Deployment stretches across the full NVIDIA hardware ladder, from RTX PCs, DGX Spark, DGX Station and Jetson boards up through RTX PRO workstations, data centers and cloud, which lets privacy-constrained workloads stay on premises without giving up the model.
Switchyard: Routing as a Cost Lever
The router targets a problem every team running agents eventually hits. Sending every call to a frontier model burns budget on trivial steps, while hand-rolling routing logic turns into integration work that delays shipping. Switchyard plugs into existing agent tooling, dispatches each prompt automatically, and exposes tunable algorithms so developers can weight quality, latency or cost however their workload demands.
NVIDIA's own numbers put task completion cost at roughly a third of running Opus 4.8 throughout, with accuracy held near frontier levels. The partner data is more interesting because the tradeoffs stay visible: LangChain cut spending 74 percent across 145 multi-turn Deep Agents tasks while sending only 7 percent of calls to a frontier model, and paid 6 percent accuracy for the privilege. Ramp held frontier-equivalent performance on its internal SWE-Bench at 58 percent lower cost and 33 percent shorter runtime, Cognition trimmed mean cost 28 percent after wiring the staged router into Devin Desktop, and Classmethod measured 27 percent savings in early internal testing.
Boomi's results speak to reliability rather than price, hitting perfect domain-routing accuracy while pushing 59 percent of traffic to a fine-tuned model five times faster and shaving 21 percent off later-turn latency, while Cadence reported a 9.9 percent efficiency gain on formal verification through its ChipStack agent. Distribution is following the benchmarks: Kong routes through Switchyard natively in its AI Gateway, LiteLLM is adding it as a proxy plug-in, and Nous Research has embedded it inside Hermes.
The Underlying Bet
Giving away both a capable small model and the logic governing when to reach past it is not charity. If ensembles become the default shape of production agents, NVIDIA has quietly positioned itself at every layer of that stack, and all of it still runs on its silicon. Lightning is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NIM microservice, with Switchyard on GitHub and partner platform support to follow.
