# Hunch — Antares Labs

Research preview · September 2026

Hunch is a decision model. Given a state in text or JSON, a question, and described candidate answers, it returns probabilities over those candidates in a single forward pass. It does not generate answer text or explanations.

## Interface

- Choice: probabilities over 2–255 supplied candidates.
- Boolean: probability of a proposition being true; the caller chooses the threshold.
- Score: distribution over 2–10 described ordered levels; the caller can compute the expected level.

These are interface capabilities, not uniform quality guarantees across all tasks and candidate counts.

## Current research preview

The current model card describes a 1.7B model adapted from Qwen3-1.7B for routing and triage. Applications under study include support intent classification, evidence verification, paraphrase detection, and selected text attributes.

## Limitations

The preview is not production-validated. Its current model card documents generalization and out-of-family calibration limitations. Evaluate the model on the task and data where it will be used. Set application thresholds deliberately, and route appropriate cases to human review.

This brief contains no benchmark, latency, general availability, or production-readiness claim. Training and release evaluation are ongoing. Downloadable weights and final release artifacts will be linked when published.

## Release direction

The project is working toward open-weight releases with reproducible data recipes, evaluation tools, and model cards. Antares Labs is the parent research identity; Hunch is its first model family.

Source: the local Hunch project model card and interface documentation, reviewed 20 September 2026. This is an introductory model brief, not a peer-reviewed paper.