Valen opens a multimodal System One preview with vision (479★)
On 23 September 2026, Valen-Team / Liuziyu77 published Valen and the Hugging Face checkpoint Valen-Preview-0923: a multimodal System One-style model that scores candidate answers over text, images and video without generating answer tokens. About 479 GitHub stars by the 29 Sep radar check.
The Preview is the 2B Sokoban RLCD line (frozen Qwen3.5-2B base required). Author demos claim ~122 ms per game step and 87.6% single-step Sokoban accuracy — UnverifiedClaim, with an outcome-conditioned game set called out on the card. Community serving: vllm-jev serve Valen-Team/Valen-Preview-0923.
Catalog entry: /models/valen. Nearby: Jev-Omni, openjev, runtime vLLM Jev.