ModelsCatalog not published
No models are published yet
The catalog is served from admin-managed data, and none is seeded. Listing models we cannot route to would be the fastest way to make this page untrue.
Model id grammar
[~]{author}/{slug}[:{variant}]author/slug is canonical and pinned. A ~ prefix is a moving alias to the newest flagship in a family — convenient, and a footgun, so the concrete id it resolved to is always reported back in response.model.
supply/auto hands the choice to triage. Anything else you name is the model you get.
Routing variants
sugar over the provider object- :nitro
- Sort by throughput; make priority-tier endpoints eligible
- :floor
- Sort by price; make flex-tier endpoints eligible
- :free
- Only zero-cost endpoints; fail if none
- :online
- Attach web search
- :zdr
- Only zero-data-retention endpoints
Adapters in the repository
| Package | Surface | What it covers |
|---|---|---|
| @supply/adapter-openai-compat | OpenAI Chat Completions dialect | Quirk profiles for openai, xai, groq and google-openai, declared as data |
| @supply/adapter-anthropic | Anthropic Messages | Request, stream, usage and error mapping onto the one internal shape |
| @supply/adapter-fal-queue | fal.ai queue aggregator | Async job polling with backoff, artifact handling and per-job cost |
The capability matrix
Quality is measured per model per task class, and every metric carries its source, method, date, sample size and model version. A cell without provenance is unknown, and an unknown cell never influences routing — the model stays fully available to anyone who asks for it by name.
- chitchat
- simple_qa
- classification
- extraction
- summarization
- translation
- rewriting
- structured_output
- code_generation
- code_review
- math_reasoning
- long_context_reasoning
- agentic_tool_use
- creative_writing
- vision_ocr
- vision_reasoning
- safety_sensitive