Thirty billion parameters is the small release this week. NVIDIA's more consequential move sits one layer up, in the software that decides which model touches each request before the model itself does any work.

One Model Among Several, By Design

NVIDIA expanded its Nemotron 3 family on August 11 with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model built for the narrow, repeated tasks that always-on agents run all day (NVIDIA, 2026). The company frames it as one piece inside a system of models. A frontier reasoning model, Nemotron 3 Ultra or a comparable model such as GPT-5.6, plans and orchestrates a workflow, while Lightning executes specific steps inside it: code review, tool calls, security alert triage, billing questions (NVIDIA, 2026).

Lightning delivers up to four times faster output and 30% faster agentic task completion than other models in its size class, by NVIDIA's own PinchBench results (NVIDIA, 2026). CrowdStrike is customizing it for cybersecurity workloads. Harvey is pairing it with its Trajectory product for legal research. CodeRabbit built a routing layer on top of it with Baseten for code review (NVIDIA, 2026).

The model ships open. Enterprises can post-train it on their own data through NVIDIA NeMo and run it locally, on NVIDIA RTX workstations, NVIDIA DGX Spark systems, or edge devices they already own (NVIDIA, 2026).

The Router Ships With Its Own Adoption Numbers

NeMo Switchyard is the second release from the same day: an open source library that routes each step of an agent workflow to whichever model handles it best, tuned for cost, latency, or quality depending on what an enterprise sets (NVIDIA, 2026). NVIDIA's internal benchmarks show it holding frontier-level accuracy while cutting task completion cost to close to a third of running Opus 4.8 alone (NVIDIA, 2026).

NVIDIA's GitHub repository still marks Switchyard pre-alpha, a status sitting next to five partners already reporting production or near-production numbers. That gap between the label and the results is worth factoring into any deployment timeline.

Several partners published their own numbers alongside the release:

• Boomi hit 100% domain-routing accuracy across five routing tests, sent 59% of traffic to a fine-tuned model running five times faster, and cut later-turn latency by 21% (Boomi, 2026).
• LangChain cut cost 74% across 145 multi-turn Deep Agents tasks by sending only 7% of calls to a frontier model, at a 6% accuracy tradeoff (LangChain, 2026).
• Ramp matched frontier model performance on its SWE-Bench suite while cutting cost 58% and runtime 33% (Ramp, 2026).
• Cognition folded a staged router into Devin Desktop, holding near-frontier performance on FrontierCode Main while cutting mean cost 28% (Cognition, 2026).
• Classmethod reported an early 27% cost reduction running opencode and Fireworks workloads through the router (Classmethod, 2026).

Every figure above comes from a named company running production or near-production workloads.

A Router Is Harder to Rip Out Than a Model

Kong is shipping Switchyard natively inside Kong AI Gateway. LiteLLM is adding it as a plugin to its proxy layer, the piece that sits between an application and every model call it makes. LangChain built its own routing benchmark on top of it.

Swapping a model checkpoint changes an API endpoint and a few configuration lines. Swapping a routing library means re-tuning the cost, latency, and quality tradeoffs across every agent workflow built on top of it, then reworking whatever gateway or proxy the enterprise wired it into.

Swapping a model checkpoint changes an endpoint. Swapping a routing library means re-tuning cost and latency across every workflow built on top of it.

Ramp, Cognition, and Boomi adopted more than a model this week. They adopted the layer that decides which model runs next, every time, for every agent they operate. That is a genuine efficiency win today, and a routing dependency an enterprise will carry longer than the model that shipped alongside it.

What Happens Next

Does NeMo Switchyard's routing logic stay open enough that an enterprise can swap NVIDIA's default algorithm without losing the Kong and LiteLLM integrations already built around it?

Does the router's default configuration route evenly among competing frontier models over time, or does cost tuning start to favor Nemotron placements as usage scales?

Boomi and Ramp published exact percentages this week. Do the next round of adopters match those numbers with the same specificity, or do vaguer results start showing up in six months?