Clef playground

The 27B precision model, live. higher-precision decisions; needs about 85 GB of VRAM to self-host.

Clef playground community-hosted

Community-hosted demo of the 27B model. Larger, slower, more precise.

Parameters
27B

Base: Qwen family

Context
24,576

Tokens read by this API (model card: 64K tokens)

Local VRAM
85 GB

At single concurrency, model-card context

What you send, and what comes back

A request carries a state and a schema of typed questions. The state is the situation the model reasons about — a support ticket, a JSON payload, a product description, or an image, since both models ship with a vision encoder. This API reads 24,576 tokens of state (120,000 characters); the published model card is wider, and anything past the served window is refused rather than silently cut. Each question declares the answers you are willing to accept.

What comes back is a probability for every allowed answer plus the winner, so a system can act on a threshold instead of trusting a single label. Nothing is generated as prose, which means there is no parsing step and no way for the model to answer outside the schema you declared.

Choosing between the two models

Pick Clef when the decision is genuinely hard — ambiguous states, subtle distinctions, cases where a wrong answer is expensive and the latency budget is seconds rather than milliseconds. Batch scoring, offline review and complex routing all fit. For interactive work, start on the smaller model instead.

Frequently asked questions

Is this the official playground?

No. It embeds a community-hosted demo. Clef and Clef-Flash are open-source models from Cloudflare, and this site is independent of them.

Does the demo cost anything?

No. The embedded demo is free, and this site’s own API has a free daily allowance counted per browser session.

Can I use the output commercially?

The models are Apache 2.0, so the weights and their outputs are usable commercially. Check the licence for the exact terms.

Clef playground community-hosted

Community-hosted demo of the 27B model. Larger, slower, more precise.

The frames load lazily. Community Spaces can take a while to wake from sleep — open one in its own tab if it looks blank at first.