Liuziyu77/Valen

What it solves

Valen is a multimodal decision model designed to provide software with a structured decision interface. Instead of generating text tokens, it evaluates text, images, and video against task instructions and outputs probabilities over a set of provided candidates, allowing for faster and more structured decision-making in real-time applications like gaming or visual perception tasks.

How it works

Built on a Qwen3.5 (0.8B or 2B) backbone, Valen uses a shared decision head to score candidates directly. This approach avoids the token-by-token generation process typical of LLMs, significantly reducing latency. It supports Supervised Fine-Tuning (SFT) and experimental Reinforcement Learning from Contrastive Divergence (RLCD) for training on custom datasets.

Who it’s for

It is intended for developers building AI agents or software that requires rapid, multimodal visual perception and structured decision-making without the overhead of generative AI.

Highlights

• Multimodal Input: Processes text, images, and video.
• Low Latency: Outperforms generative models in decision speed, as seen in its ability to solve puzzles (e.g., Sokoban) in milliseconds per step.
• Structured Output: Returns decision probabilities rather than free-form text.
• Customizable: Includes tools for data processing, SFT, and RLCD training to adapt the model to specific tasks.