A decision model returns an answer, not a paragraph
Decision models are the quiet half of production AI: the part that labels, routes, scores and approves. Clef and Clef-Flash are Cloudflare’s open-weight entry into that category.
The definition, without the marketing
A decision model is trained to map a state and a schema of typed questions onto a distribution over the allowed answers. You give it the situation and the list of answers you are willing to accept; it gives you back a probability for each one. There is no free text, so there is no parser, and no prompt that can be argued out of the schema.
That is a different contract from a language model. A language model is optimised to continue text plausibly. A decision model is optimised to be right about a bounded choice, quickly, and to say how sure it is.
- Input: a state (text, JSON, image, video frames) plus typed questions with allowed answers.
- Output: a probability per allowed answer, in one forward pass.
- No chat, no tool calls, no hidden reasoning to parse.
Why this matters in production
Most real systems do not need a paragraph. They need a routing decision, a severity label, a yes/no with a confidence number, or a moderation verdict that has to be defensible after the fact. When you get those from a generative model you pay for tokens you throw away and you inherit a parsing failure mode.
A decision model removes that class of bug by construction: the answer set is fixed, so the failure mode is “low confidence on the wrong option” rather than “the model replied with an essay and the regex missed”.
Where decision models fit next to LLMs
They are complements, not replacements. The usual production pattern is a language model that understands an open-ended request, and a decision model behind it for the bounded choices — which tool to call, whether to escalate, which queue a ticket belongs in.
Clef-Flash is fast enough (about 39 ms median) to sit inline in that loop, which is what makes the pattern practical rather than theoretical.
Start with a state and one question
The playground on this site lets you paste a state, write one or two typed questions, and see the probability distribution immediately. That is usually enough to tell whether your problem is a decision problem or a writing problem.
Frequently asked questions
What is a decision model in one sentence?
A model that takes a situation and a fixed set of allowed answers and returns a probability for each answer, instead of generating text.
Do decision models replace LLMs?
No. They replace the fragile part of an LLM pipeline — the step where you try to extract a bounded decision from generated prose.
What are typical use cases?
Routing an agent to a tool, triaging support tickets, labelling images, scoring leads, moderating content, and any gate that has to decide before a human sees the queue.
Keep reading
Model notes
One short email when a new decision model lands or a benchmark changes. No spam, unsubscribe in one click.