Act on this, mostly facts

Jeffan-31B Gemma fine-tune ranks first on Jev Decision Index

What to do

Test Jeffan-31b on vLLM 0.29.0 if you need a single-pass decision model, keeping inputs under 8,192 tokens.

Why

The Apache 2.0 model scored 69.8 and ranked first as of October 21, which is verified.

What the research found

Jeffan-31b, a Gemma-4-31B-it fine-tune under Apache 2.0, scored 69.8 and ranked #1 on the Jev Decision Index as of 21 Oct 2026; it is a single-forward-pass decision model verified on vLLM 0.29.0 with an 8,192 token input limit.

Claims and how well the evidence supports them

Jeffan-31b scored 69.8 and ranked #1 on the Jev Decision Index as of 21 Oct 2026 97%
The model is built on google/gemma-4-31B-it under Apache 2.0 98%
Verified with vLLM 0.29.0 on one NVIDIA B200 in bf16 99%
Inputs are limited to 8,192 tokens 98%

Evidence

huggingface.co https://huggingface.co/tattrongvu/Jeffan-31b-1710

Actionability rated medium.