Models
What you can call, how to name it, and why a model might be ready, warming, or offline.
The catalog
The set of callable models is curated centrally and shown in the marketplace. You reference a model by its slug — the same string you pass as model. Slugs encode the family, size, and quantization, for example llama-3.2-3b-instruct-q4f32_1.
List the live, authoritative catalog programmatically:
1curl https://mesh-zeta-eight.vercel.app/v1/models \2-H "Authorization: Bearer $PIEDPIPER_API_KEY"
This returns the full enabled catalog in OpenAI shape — including models that no host is currently serving, so you can discover everything Pied Piper knows about, not just what's warm this second.
Availability & warmth
A model in the catalog is only callable when some live host is serving it. Hosts advertise models in one of two states:
- Warm (
ready) — loaded in GPU memory. Requests start generating immediately. - Loadable — advertised but not loaded. The first request triggers a warm-up; Pied Piper holds the request (up to ~60s) rather than failing.
- Offline — in the catalog but no live host. You get a
503 service_unavailable_errorif you call it.
X-Pied-Piper-Host to avoid the cold-start wait — see routing headers.Browser-tier vs native-tier
Pied Piper has two classes of host, and that shapes which models appear. Browser and extension hosts run quantized small-to-mid models through WebGPU; native nodes can pool multiple machines to serve large, sharded models (e.g. a frontier DeepSeek checkpoint split across GPUs via llama.cpp RPC). As a borrower you don't choose the tier — you choose a slug, and dispatch finds a host that can serve it.
Pinning to a rented session
If you hold a rental token for a specific session, pass it as X-Pied-Piper-Provider on GET /v1/models to narrow the list to exactly the models that session has loaded. Without it you see the whole catalog.