qev-0.8b: New State-of-the-Art Tiny Local Decision Model
AustinFu has introduced qev-0.8b, the smallest model in the Qev decision model series featuring a 0.8-billion parameter size. It is designed to take context, a question, and candidate answers to output a specific choice along with probabilities for every option. This model provides an efficient option for edge computing and local execution where users need typed answers and calibrated confidence instead of free-form text. It shares the same API as larger variants like Qev-2B, Qev-4B, and Qev-9B, offering flexibility for developers building resource-constrained local applications. The model utilizes the Qev v0.3.0 Snake continuation checkpoint, which incorporates Snake supervision and replay of original general decision tasks. It supports multiple formats including Choice, Noul (yes/no), and Score (ordered ratings) through a unified API.
## BACKGROUND
Unlike traditional generative language models that produce free-form text token by token, decision models are specialized AI systems designed to process prompts and discrete candidate choices simultaneously in a single forward pass. They return structured, typed answers with calibrated confidence scores, making them ideal for classification and decision-making tasks.