Valen
Apache-2.0 multimodal System One models (0.8B, 2B, 4B) that take text, images or video and return probabilities over supplied candidates.
Overview
Valen brings visual perception to Jev-style decision-making. Inspired by Jev, it evaluates text, images and video against task instructions and returns probabilities over candidates the caller supplies. The repository includes the model implementation, data processing, supervised fine-tuning and experimental RLCD training, inference and evaluation, with support for training on your own data.
Checkpoints at 0.8B, 2B and 4B are on Hugging Face, along with ONNX and GGUF versions, a 100k training dataset, the VisualDecisionBench benchmark and an online demo; API access is also offered. The README reports VisualDecisionBench-Image accuracy rising from about 76 to 83 percent and video from about 79 to 87 percent across the three sizes, compared with published baselines and on JevBench. Models are built on Qwen3.5 with a mixer and shared-state inference.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which supplied candidate best matches this image or video for the instruction? | — |
How it works
- Provide text, an image or video, an instruction and candidates.
- Valen runs shared-state inference.
- It returns probabilities over the candidates.
- Optionally fine-tune on your own data.
Strengths
- Multimodal decisions, which Jev itself does not support.
- Open weights, training code and data.
- Multiple sizes and ONNX/GGUF formats.
Caveats
- Benchmarks are largely the authors' own (VisualDecisionBench).
- RLCD training is marked experimental.
At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch, ONNX, GGUF, hosted API
- Has code
- yes
- License
- Apache-2.0
- Stars
- 790
Use cases and integrations
How curated lists describe it
Apache-2.0 multimodal System One models (0.8B, 2B, 4B): text, images, and video in, probabilities over the supplied candidates out, with training code, training data, and an online demo
AnotiaWang/awesome-decision-modelsTrain a Jev-like multimodal model by yourself. System One Model, now with vision.
wh000wh000/awesome-jev-liveTrain a Jev-like multimodal model by yourself. System One Model, now with vision.
hellogumbo/awesome-jevMultimodal System One model, text and images and video in and decision probabilities out, with training and evaluation code and weights on Hugging Face.
MrJev/awesome-jev
Frequently asked questions
What inputs does Valen accept?
Text, images and video.
Which sizes exist?
0.8B, 2B and 4B.
Can I train my own?
Yes, training code and data are included.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPythonIndependent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPython