Qwen3.5 4B + grabbing logits is almost "Jev"? Or even just Qwen Reranker?

A developer community discussion explores the capabilities of Qwen models in relation to the "Jev" model, which outputs probabilities for given choices. The user initially considered if Qwen reranker models could replicate this functionality. However, they found even better results by implementing Qwen 4B and assigning each option to a letter to grab the logit probabilities. This suggests that Qwen 4B, by extracting logit probabilities, can achieve similar or superior outcomes to the Jev model.