Open-source Valen: a visual decision model scoring candidates in 122–128 ms

aigclink · x · 2026-09-28

Valen extends System One decision-making to vision: it takes text, images or video and outputs probabilities over given candidate actions instead of generating tokens. Built on a Qwen3.5-0.8B/2B backbone with a shared decision head, it targets high-frequency agent reflexes like GUI popup detection, game movement and robot obstacle avoidance. Official numbers: 122–128 ms per decision, four parallel rounds in 1.24 s. The repo includes implementation, training and evaluation, with support for training on your own data.

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

Original post →

More from coding & agent

coding & agent channel →