NVIDIA: Switchyard

NVIDIA: Switchyard is an open-source model router from NVIDIA. Rather than being a single model in the usual sense, it is a routing layer that switches between multiple models, and its stated purpose is to optimize the cost of requests. By default it uses market data to select the most popular option among the models available to it. It sits in the text category and is identified on api14 by the model id nvidia/switchyard.

As a router, Switchyard takes text in and returns text out. It accepts text input only and produces text output only, so there is no image, audio or file handling to plan around. Its context window is 1,000,000 tokens, which is large enough to hold very long documents, extended conversation histories or sizable codebases in a single request. The provider description does not specify a maximum output length, and no supported request parameters are listed for this model.

The practical case for Switchyard follows from what it is: a cost-aware dispatcher. If you run a workload where many requests do not need your most expensive model, routing them through Switchyard lets the router pick among models based on market data instead of you hardcoding a choice per request. The million-token context window also makes it viable for long-input tasks such as summarizing lengthy reports, searching across large document sets, or maintaining long-running chat sessions where you would rather not manage truncation yourself. Because the output is plain text, it fits standard generation, extraction and question answering flows.

You call it through api14's OpenAI-compatible API using the model id nvidia/switchyard. Point your existing OpenAI client at the api14 base URL and pass that id in the model field of a chat completion request. No supported parameters are listed for this model, so treat it as accepting the basic request shape (model plus messages) and avoid assuming that optional knobs such as temperature, top_p or max tokens will be honored. If you need a specific parameter to take effect, verify it against the live behavior before relying on it in production.

One thing to weigh is that Switchyard is a router, not a fixed model. The underlying model that handles a given request can vary, which means output style, latency and quality may differ between calls in ways you cannot fully predict from the model id alone. If you need deterministic, reproducible behavior from one known model, a direct model id is likely a better fit. Switchyard is most useful when cost optimization across a mixed workload matters more than pinning every request to a single model.