Cloudflare: Clef

Cloudflare: Clef is an open-source multimodal decision model from Cloudflare, released on 1 October 2026. It is a fine-tune of Qwen3.8-27B and is served on Workers AI. Its main category is decisions, and that single category describes what the model is built to do rather than treating it as a general chat model.

Clef accepts text and image input. Its output modality is decisions, so instead of free-form prose it returns answers to a schema of typed questions. The context window is 65,536 tokens, and the maximum output is 14,745 tokens. The provider description frames the input as a state, given as text or structured JSON, paired with a schema of typed questions. That combination of state plus schema is the core of how the model operates.

The practical use cases follow from that design. If you have a state you can express as text or JSON, and a set of typed questions you want answered about it, Clef is built for that shape of task. Examples include classifying or routing incoming items, extracting structured fields from a document, and making a choice among defined options. Because images are accepted as input, the state can include visual material alongside text, which suits tasks where a decision depends on something that is easier to show than to describe. The typed-question schema means the answers arrive in a form your code can consume directly, rather than requiring you to parse prose.

You call Clef through api14's OpenAI-compatible API using the model id "cloudflare/clef". The model lists no supported parameters, so you should not expect to pass sampling controls or other optional request fields and have them take effect. Send your state and your schema of typed questions as the input, and read the decisions from the response. Because the output is decisions rather than text, plan your integration around a structured result rather than a message string.

The context window of 65,536 tokens and the output ceiling of 14,745 tokens are the main practical bounds to keep in mind. If your task is open-ended generation, conversation, or anything that needs free-form text output, a model whose output modality is text will fit better, since Clef returns decisions. If your task is a decision over a state you can supply as text, JSON, or an image, and you can express what you want to know as typed questions, Clef matches that shape directly.