Nvidia Switchyard：模型路由降低66% Agent成本

enterprises that run always‑on AI assistants are stuck in a classic cost dilemma. Sending every request to a cutting‑edge foundation model guarantees quality, but the bill climbs astronomically. Building custom routing logic to divert simple queries to cheaper models solves the price problem—until the workflow changes and the routing code must be rewritten. Nvidia aims to eliminate that double‑handed headache with a two‑pronged solution released on Tuesday.

How Switchyard Redefines Model Routing

Switchyard is an open‑source Python library that sits at the heart of an AI agent's execution graph. Instead of hard‑coding which model handles each step, developers declare a set of candidate models and let Switchyard pick the best fit in real time. The decision engine evaluates factors such as token count, latency budget, and estimated inference cost, then routes the request to the most efficient model for that specific sub‑task.

The key innovation is the "mid‑task reshuffle." When an agent progresses through a multi‑step workflow—say, extracting entities, summarizing text, and generating a response—Switchyard can switch models between steps. A lightweight expert handles the extraction, a medium‑sized model tackles summarization, and a flagship model only runs for the final creative output. Nvidia's internal tests showed that this dynamic hand‑off reduced overall compute spend by about 66 percent compared with a single‑model pipeline.

Nemotron 3.5 Lightning: A Specialist Mixture‑of‑Experts Model

Alongside Switchyard, Nvidia introduced Nemotron 3.5 Lightning, a 30‑billion‑parameter mixture‑of‑experts (MoE) architecture designed for high‑volume, specialized agent workloads. Unlike monolithic models, the MoE design activates only a subset of its experts for any given input, dramatically lowering inference cost while preserving the expressive power of a much larger network.

Lightning is trained on a curated mix of code, documentation, and conversational data, making it especially adept at routine developer‑support queries, ticket triage, and other repetitive agent tasks. Because it is open‑source, companies can fine‑tune the model on proprietary data without licensing hurdles, and they can embed it directly into the Switchyard routing table as a low‑cost default option.

Real‑World Impact and Future Outlook

In Nvidia's benchmark suite, a typical enterprise AI workflow that previously required a single 175‑billion‑parameter model was re‑engineered with Switchyard and Lightning. The total token‑processing cost dropped from $0.12 per request to $0.04, while latency remained within service‑level agreements. For large organizations that process millions of requests daily, the savings translate into multi‑million‑dollar reductions.

Beyond cost, the approach simplifies maintenance. Teams no longer need to rewrite routing scripts whenever a new model is added or a workflow is altered; they simply update the Switchyard configuration file. This decouples model management from business logic, a win for DevOps and MLOps alike.

Analysts expect the combination of dynamic routing and MoE specialists to become a standard pattern for AI‑first products. As more vendors release open‑source Mixture‑of‑Experts models, Switchyard's model‑agnostic design positions it as a universal orchestrator for heterogeneous AI stacks.

Takeaway: Nvidia's Switchyard library and Nemotron 3.5 Lightning give enterprises a plug‑and‑play way to match each AI sub‑task with the most cost‑effective model, delivering up to a two‑thirds reduction in per‑task spend while keeping performance on target.