Inference Model Catalog

This catalog lists the models served on the inference API: the models you can call at api.cyan.kiwi/v1 with an API key. It is the “model catalog” the Terms of Service refer to. It is not a list of the open-weight models cyankiwi publishes for download on Hugging Face. Those are a separate surface with their own page at /models and their own terms (clause 12 of the Terms), and a model appearing on one list says nothing about whether it appears on the other.

The machine-readable catalog is GET https://api.cyan.kiwi/v1/models: public, no key needed. This page is that same catalog rendered readably: it is fetched live from the endpoint, so what it shows is what is deployed, never a hand-maintained copy that can go stale.

Clause 6 of the Terms of Service makes your use of each model, and of its outputs, subject to that model’s upstream licence as identified on this catalog, including any operative condition shown on the entry. The licence, a link to its text, and any condition that falls on you are published per model, here and in the license field of the endpoint.

Loading the live model catalog…

What you are served

Where an entry is marked quantized by cyankiwi, you are served our build, not the publisher’s released weights. The precision we serve at travels with the model’s name on each entry, because it is part of what the model is. An entry showing fp16, bf16 or fp32 is served at the publisher’s own release format.

Quantization is lossy. A quantized build is a compression of the published model, not a copy of it, and its output can differ from the publisher’s original weights. The upstream name on an entry tells you which model we started from; it is not a representation that what we serve is equivalent to, or a substitute for, that model at full precision. Evaluate a model on your own workload before you depend on it. Clause 12 of the Terms takes the same position for the models we publish for download.

Licences

Each entry’s licence is recorded at intake by reading the upstream repository’s own licence file, never the hub tag, and the licence URL is pinned to a fixed upstream revision, so the text you are sent to is the text that was read. A model with no recorded licence is not listed on this catalog and is not served. The record is made at a point in time: it is not a continuous audit, and publishers can amend their terms after release. If a licence matters to your use, read the linked text. The intake practice is described at the Trust Center.

Conditions mean what they say. Where an entry shows a condition, it is an obligation the upstream licence puts on you rather than on us. Where an entry shows none, we identified none falling on you as an API user, not that the licence is silent on everything. If your use goes beyond calling the API, read the linked licence.

Capabilities, preview and prices

Declared capabilities record a test, not a guarantee. supported_features, supported_sampling_parameters and the modalities on the endpoint are the ones we validated end to end against that model. Anything not declared is untested and unsupported; the API is a transparent relay and will not necessarily reject undeclared content. Verify a capability by its effect before you depend on it.

Preview. A model whose entry carries is_ready: false is a preview model: no availability or retirement commitment, may change or be withdrawn at any time (Terms, clause 6). An entry without that flag is generally available.

Prices are USD per token on the endpoint (exact decimal strings), shown here per million tokens, and exclude VAT and any equivalent sales tax (Terms, clause 15). Cached input tokens bill at the cache-read rate when the engine reports a cache hit: a discount when available, never a guarantee. Price changes apply to requests served after publication (Terms, clause 15).

Where inference runs

Inference runs on engines selected and configured by cyankiwi, on serverless GPU infrastructure. The canonical statement of where it runs (the regions, and the prompt cache we operate) is the sub-processor register. We do not currently offer a per-customer choice of region, and using a model does not route your requests to its publisher.

Retirement

We may add, change or retire models at any time; retirements of generally available models are announced on this catalog, and clause 6 of the Terms of Service is the governing text. Preview models may disappear without notice, and we do not commit to serving any future version of a model.

Names and marks

Model and publisher names on this page identify which upstream model an entry is built from; they are the property of their respective owners. cyankiwi is not affiliated with, endorsed by, sponsored by or authorised by any model publisher, and what we serve is not supplied, warranted or supported by the model’s publisher.