Valen
Text, images, and video in. Decision probabilities out.
Valen (万澜) brings visual perception to System One decision-making. Inspired by Jev, it evaluates a shared state against task instructions and returns probabilities over supplied candidates.
The repository includes the model implementation, data processing, SFT and experimental RLCD training, inference, and evaluation tools.
Choose a model
| Model | Weights |
|---|---|
| Valen 0.8B | Valen-0.8B |
| Valen 2B | Valen-2B |
| Valen 4B | Valen-4B |
Released formats and earlier checkpoints are listed in the Hugging Face organization.
Three decision types
- Choice: distribute probability across named options.
- Noul: estimate whether a statement is true or false.
- Score: assign probability to ordered grades.
The same state can be shared by multiple questions. Read the data format and shared-state execution guide for the full semantics.
Get started
- Serve with SGLang or vLLM, use Transformers locally, or follow the hosted API quickstart.
- Explore image and video inputs.
- Read the architecture and configuration reference.
- Evaluate a checkpoint or compare models with JevalKit.
Source repository · Cookbook · Try the demo
The hosted API names true/false questions predicate and returns probability; local inference and evaluation retain the noul naming. Use the API reference for hosted request and response fields.