Valen on GitHub: train a Jev-like visual decision model yourself

aigclink · x · 2026-09-28

The GitHub repo for Valen (Liuziyu77/Valen, 390 stars) is now public. Inspired by Jev, it ingests text, images and video and returns probabilities over supplied candidates via a Qwen3.5-0.8B/2B backbone plus a shared decision head — no answer tokens generated. The repo ships model implementation, data processing, SFT and experimental RLCD training, inference and evaluation commands, and supports training on your own data. An online demo (Valen-Preview-0923) shows it reading board images and picking moves against a 27B generation model.

Related event: Open-Source Visual Decision Model Valen Launches with 122ms Decisions(2 posts)→

Original post →

More from coding & agent

coding & agent channel →