Jev 替代模型 S1 榜单：Autojev 27B 等五个开源选项

Cuth @ItsCuthulhu原推发布于 2026.09.23

在 X 上打开原推

UPDATED S1 BENCH RANKINGS - (Top 5 Jev Alternatives) As one of the leading Jev experts of this generation - I would say that any one in the image are good alternatives for Jev, but be mindful of Model speed. Simplejev-qwen3.8-flash-next, litjev-27b and AutoJev 27B + voting are not worth it. Some of these were created by me, so you won't find them anyway. Just be mindful that not every option a worth alternative. __ Here are the new top 5 Jev alternatives: 1. Autojev 27B: The best Jev alternative by far. @denisyarats created it (CTO of Perplexity). It is both faster and more accurate than Jev... But it is not by much. If you want a great open source Jev, then this is your pick. It is however 27B, so you may need a DGX Spark or something similar. 2. SimpleJev Qwen 27B: Building on the classic Qwen 3.8 27B, this model will also fit on your DGX Spark or similar but it doesn't match Autojev 27B. Without an H100, you will probably not get it to match it either. In this case, it is best to use AutoJev 27B if you can. Notably, you can use Simple Jev for any model: https://simple-jev.featherless.ai/ 3. djev - This model seems to be a variant of 'Diffusion Gemma as Jev'... It performs very well but not as well as Jev. The benefit is the tokens that it can read in sit around 4096 and it is slightly faster than Jev itself. https://github.com/mmastrac/djev-spark 4. Decider 35B-A3B + per-subset prompts - This is a slight variant I made. It scores about 74% accuracy, which is right under Jev... The primary benefit is that it registers as ~3x faster than Jev while losing only ~3% accuracy. This is my System One model of choice because the decreased accuracy is made up for by massive speed increases. I can provide the variants on Github with links here: http://bench.jakecuth.com It's origins are from this Github: https://github.com/Mapika/decider 5. Lastly, I would say Reflex-4B or Kev-4B, which I recommend for anybody trying to use Jev locally without a DGX Spark or a lot of GPU/RAM space. These models are tied because they get similar accuracy and both are wicked fast. At some point, size and speed matters more than accuracy so for the edge cases, these are great alternatives. If you have a new submission... Leave it in the comments and my agents will pick it up. Hope to see new contenders soon!