# fez Fez is an experimental decision-model platform that turns workflow context into probabilities for model routing, tool selection, escalation, and support triage. It helps teams train and evaluate specialized models for repeated decisions in AI systems. - Findings24: - Site: - Preços: Freemium - Categorias: Gateways e roteamento de IA, Avaliação de LLMs ## O que é fez? Fez is an experimental platform for training and evaluating specialized decision models for business workflows. Its core interface is **context in, probabilities out**: instead of generating a full response, a model estimates the available actions or outcomes. This is intended for repeated decisions such as model routing, tool selection, escalation, and support triage. The documented workflow uses authorized labeled examples, acceptance criteria, training jobs, and evaluation results. Fez also provides a playground where users can supply text or JSON context, define choice, yes/no, or score decisions, and inspect the probability assigned to each option. Fez 0.8B is described as an experimental model. Published research includes benchmark and training-to-testnet evidence, but the site says the measurements are diagnostic, hosted cost has not been measured, and the checkpoint comparison does not establish a production SLA. Customer delivery is planned; the training dashboard is currently a development preview. ## O que fez pode fazer? ### Structured probability outputs Fez returns probabilities for defined decisions rather than a generated text answer. This supports downstream logic that needs a choice, score, or confidence distribution. ### Model routing decisions The product can evaluate whether a request should remain with a smaller model or be sent to a larger model, reserving larger-model calls for cases that need them. ### Tool and action selection Given agent context, Fez can score candidate tools or next actions without requiring a full language-model response for every selection. ### Training and evaluation workflow The development preview is designed for labeled datasets, acceptance criteria, training jobs, and inspection of evaluated model artifacts. ### Decision playground Users can test text or JSON context and define decisions as choices, yes/no questions, or scores. The playground displays the probability of every option and sends requests to the connected Fez checkpoint. ### Experimental research evidence The research area records benchmark metrics including accuracy, Brier loss, calibration-related measurements, confidence mistakes, and localhost latency, with stated methodological limitations. ## Casos de uso ### Route requests between models An AI application can classify whether a request is short and well specified enough for a smaller model or requires a larger model for ambiguity, long context, or multi-step reasoning. ### Choose the next agent tool An agent can use the decision output to select among available tools, such as looking up an order, searching documents, or handing a case off, without generating prose for the selection step. ### Escalate uncertain workflows Teams can label outcomes to assess whether an agent should continue, retry, stop, or ask a person for review before taking its next action. ### Triage support requests The playground documents support-oriented decisions including routing a customer request, checking urgency, and scoring tone. These examples can be expressed as structured choices or scores. ## Perguntas frequentes ### What does Fez return? Fez is designed to return structured probabilities for defined decisions. It does not generate a text answer for each decision, and the playground can show the probability of every available option. ### How is Fez trained? The documented workflow starts with authorized labeled examples and acceptance criteria, followed by training and evaluation. The training dashboard is a development preview for submitting datasets, following jobs, and inspecting evaluated artifacts. ### Is Fez available as a production service? The site describes customer delivery as planned and labels the training interface as not yet a launched training service. The available material does not document a generally available hosted service or deployment interface. ### What model does the playground use? The playground identifies the connected model as Fez 0.8B, experimental. It says answers come from that connected Fez checkpoint and that there is no fallback model. ### Does the published research prove production performance or savings? No. The site says its benchmark measurements are diagnostic, the comparison does not establish a production SLA, hosted cost was not measured, and customer cost savings have not yet been measured. ## Informações rápidas - **Category**: AI decision model platform - **Primary users**: Teams building AI agents and business decision workflows - **Core output**: Probabilities over defined decisions - **Documented model**: Fez 0.8B, experimental - **Current status**: Early access; customer delivery planned - **Access channel**: Fez Discord community and development preview