NVIDIA Nemotron · Nebius Token Factory
Switchyard routes every AI call in your app to the right NVIDIA Nemotron model, with tests that prove it. Then it helps you build, secure, ship and keep your app running on Nebius.
LLM calls moved to Nemotron in the latest recorded run on a real repo. The rest failed their tests and stay on their current model.
cheaper per request for the switched calls than the model they use today, at Token Factory list prices.
spent on Token Factory to scan, test, evaluate, repair, patch and audit every call in that repo.
Read from the recorded run below · nothing here is estimated
One toolkit · every model step on Nemotron · six steps
Each call is tested on Nemotron 3.5 Lightning, 3 Super and 3 Ultra and moves to the cheapest one that passes, with a pull request and the evidence. switchyard router then sends each feature to its tested model.
Nemotron Super plans and writes a FastAPI app with tests, a non-root Dockerfile and a Token Factory client, then changes it on request.
switchyard new "<idea>"A comment on every pull request that touches AI code: the cost at your volume, the Nemotron model worth testing, and security issues on the changed lines.
switchyard pr-reviewLeaked keys, model output reaching eval, shell or SQL, user input in system prompts, missing timeouts and token caps. Then injection probes on Nemotron.
Free audit of any repoRuns the security checks, builds the image and deploys it to Nebius Serverless with the app's Token Factory key, then health-checks the URL.
switchyard deployOn-call for the deployed app: finds the change that broke it, proposes one fix, applies it when you approve and proves it recovered.
switchyard oncallThe app
Open a recorded migration to see every call site, the tests behind each switch, the patch and the bill. Or start your own.
Most apps send every call to one frontier model. Many of those calls are classification, extraction or short summaries that a smaller open model handles just as well. Switchyard finds each call, tests it on the NVIDIA Nemotron family, and only switches the calls that pass.
Pick a recorded run on the left to see a full migration, or start a new one.
Pick a file.
Runs on your machine in a container: non-root, bound to 127.0.0.1 only, the same image switchyard deploy ships to Nebius.