Developers introduced the Reflex model for making structured decisions

The team of developers released Reflex — a compact model based on the Qwen3.5-4B architecture, optimized for operation on GPU. The tool is designed for automated decision-making and answering structured questions with an indication of the level of confidence in the result.

The model supports multimodal input, including processing of text and images. The system classifies requests into three main categories: selection from proposed options (choice), answers to questions with uncertain answers (noul), and rating (score). The use of the 4-billion architecture allows the solution to run on consumer equipment with minimal delays.

The technical implementation of Reflex is focused on the accuracy of probabilistic indicators. The model provides not only the final answer, but also a confidence coefficient, which is critical for integration into automated pipelines that require data verification without human intervention.

The source code of the project is published in open access on GitHub. The developers position the tool as a solution for tasks that require high processing speed while maintaining logical strictness of answers. The integration of the model allows for the automation of routine data classification processes in real-time.