The 35B step up in the open decision line. Autotrust’s JEV-35B (weights October 11) is a finetune of Qwen3.5-35B-A3B: a 256-expert MoE that keeps about 3B parameters active per token, scored at 59.73 on the public Decision Index 0.3 suite, ahead of the smaller JEV siblings and aimed at the same job: read a question with up to 256 fixed options, return one label and a probability per option in a single forward pass.
Latency and fit. Median of 184 ms per decision on one B200; the same card runs the model in bf16 with room to move, and a q4 build at roughly 17.5GB of weights plus 80MB of cache per 1,000 tokens (40 layers, 2 KV heads, 256 head dim) fits a 24GB card with real context. Community quants shipped within days (mradermacher), Apache 2.0 means the finetune recipe is inspectable too.
The vision sibling is in the family. On the rebuilt vision benchmarks JEV-35B scored 71.94 against JEV-27B-VL’s 71.67: a tie within noise, so the 35B covers the vision cases without a second model.