Decision model
A decision model reads a situation and a list of questions, then returns a probability for each allowed answer instead of writing text.
A chat model generates words and leaves you to parse them. A decision model is built for the step where an agent must choose: you send the situation and typed questions, such as true or false, pick one of these teams, or rate severity from one to four. The model scores every allowed option in a single pass, so the result is a set of probabilities your code can threshold, route on or send to a human. Cloudflare's Clef, which is post-trained from Qwen, is an open example, and its request format is compatible with Typesafe AI's Jev. Liquid AI's d1, AutoTrust's Decider models, Perplexity's pplx-decider, Sarvam's Prism, which is built on a 105B model, Abacus.AI's Smaug-Decisions, which also takes audio, Hindsight's hs-decide-1-flash and Adventists AI's DuplexJev, which decides turn-taking, interruption and tone straight from speech audio, are others.
Because no text is generated, these models can be fast and consistent, and you can change the labels without retraining. They cannot explain themselves or take actions, and very reasoning-heavy questions may suit a larger chat model better. Treat the probabilities as scores to test against your own labeled examples, because calibration varies, and vendor benchmark tables come from the vendor's own runs.
Related terms
- AI agent An AI agent is a system in which a model plans and carries out multi-step tasks by calling tools, such as running code, searching the web or editing files.
- Tool calling (function calling) Tool calling lets a model ask your program to run a function, such as a search or a calculator, and then use the result in its answer.
- Multimodal model A multimodal model can handle more than one type of data, such as text, images, audio or video, as input, as output, or both.
- Fine-tuning Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain.