jevIMAGE
openjev: 0.8B Qwen replica of Jev supports image, video and text input
Someone trained an open-source version of Jev on a 0.8B Qwen model that accepts image, video and text input, is calibrated with RLCD, and runs about 5x faster than Jev with sglang.
alex nikolic
@justALEXWORTEGAArchived 2026.09.22Originally posted 2026.09.21
alex nikolic @justALEXWORTEGAOriginally posted 2026.09.21
https://huggingface.co/spaces/AlexWortega/openjev HUGE thx for @_akhaliq for domain, now openjev is: - 0.8b qwen - image-video-text train for input, audio coming дфеук - calibrated with RLCD (its somehow better then vanila grpo) - sglang support (5x faster then jev by @typesafe lol) 4b and 35b and blog coming later this week
This post quotes another post: “Training an MLP on Qwen 4B to replicate Jev”
x.com/justalexwortega/status/2102012696111575334Archived 2026.09.22 10:21