← sources

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

source: reddit (reddit:LocalLLaMA)published 2026-09-16 15:10ingested 2026-09-17 02:39

retained text

Saw the new [Introducing System One Models & Jev - TypeSafe AI Blog](https://typesafe.ai/blog/introducing-system-one-models-and-jev) Jev model which just outputs probabilities given choices and I thought it sounded a something the Qwen reranker models could do? Then I tried implementing it and got even better results with Qwen 4B + just assign each option to a letter and grab the logit probailities for each one.

Results and samples at [https://github.com/TheoLeeCJ/openjev](https://github.com/TheoLeeCJ/openjev), you can try it in your browser too at [openjev.com](http://openjev.com)

Is there something I'm missing here?

[Almost not-Jev?](https://preview.redd.it/ttzzeckjdwph1.png?width=1007&format=png&auto=webp&s=04eb2669ab2306e319892cb271505eb4bb017c5b)

[Results](https://preview.redd.it/108okswhdwph1.png?width=1047&format=png&auto=webp&s=3264fcef0006c118a03b2391ae712e571eaa3d6a)

document fetched 2026-09-17 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1whzy7j/qwen35_4b_grabbing_logits_is_almost_jev_or_even