Every size ships as native weights, ONNX and GGUF. The 0.8B, 2B and 4B models are live in the playground and the free API, with no key and no download.

What is new​

Bounding boxes. A bbox question returns a box, its centre point and a confidence. It sits in the same request as your predicate, choice and score questions, and all of them are answered in one forward pass. See bbox questions.

Block-structured attention. The shared prefix holding your text and media is bidirectional and is computed once for every question. Decision questions read it through one bidirectional block; each bbox question has its own. Read the architecture notes.

Five sizes. 0.8B, 2B, 4B, 9B and 27B. Start with 0.8B on a laptop CPU and move up when your task needs it. Which version should I choose?

Results on Jev Decision Index​

On Jev Decision Index 0.4.1, Valen-Decision-1.5-27B scores 82.9 on the public Vision track, the highest of the 83 entrants, and 74.7 on the public Text track, third of 298. Its full Vision score, which adds private test sets, is 73.3, fifth on the board. Rankings are from the 11 November 2026 snapshot.

Scores are normalized index scores from 0 to 100. Per-size results

Run it your way​

Get the models​

• Valen-Decision-1.5 collection
• Source repository
• Cookbook

Source: Valen release notes.

The release brings benchmark and model-size filters, model comparison, probability metrics, and downloadable results into a single place.

A common evaluation workflow​

JevalKit provides unified model interfaces and prepared datasets for eight Jev-related benchmark families. The published dataset collection contains ten TSV views; image and video views are subdivisions of VisualDecisionBench.

Twenty model families are registered. Registration is separate from complete evaluation coverage: the leaderboard distinguishes complete evaluations, smoke tests, and imported predictions.

Explore the results​

• Filter by benchmark and model size.
• Compare model results and probability metrics.
• Download the exported evaluation results.
• Inspect coverage and result provenance before comparing scores.

Open the leaderboard

Run an evaluation​

Use the JevalKit quickstart to install dependencies, download a model and dataset, and run an evaluation. The benchmark protocols explain how each suite is scored.

GitHub · Datasets

Source: JevalKit release notes.

ONNX and GGUF releases are available through the Valen Team Hugging Face organization. The API cookbook includes examples for the preview checkpoint and the 0.8B, 2B, and 4B models.

Use a hosted model​

The API cookbook contains executable examples with input previews and saved outputs. Start with the API quickstart for the request schema.

Run an exported model​

The published exports include a FORMAT_USAGE.md guide and a dedicated valen_formats.py inference entry point.

ONNX uses ONNX Runtime for the visual backbone, language backbone, and Mixer. GGUF retains the original tensor precision and uses a custom valen_qwen architecture, so use the accompanying Valen inference entry point.

Browse released models and formats

Export from source​

The repository includes tools to export and verify formats, and to check agreement on text, image, and video inputs across execution modes.

Read the export guide

Source: Valen release notes.

Alongside the models, we are releasing VisualDecisionBench, a benchmark for decisions grounded in images and video.

A structured decision interface​

Valen evaluates a shared state against task instructions and returns probabilities over supplied candidates. It supports Choice, Noul, and Score questions, giving software a structured interface to model decisions.

The repository includes model implementation, data processing, inference and evaluation commands, and SFT and experimental RLCD training support.

Models​

VisualDecisionBench​

The full released evaluation set contains 1,696 shared states and 9,143 questions, split into 2,000 image questions and 7,143 video questions.

Explore the benchmark · Download the dataset

Start building​

Run Valen locally or explore the source repository.

Source: Valen release notes.

Valen brings visual perception to System One decision-making: text, images, and video in; decision probabilities out.

Open from the start​

The repository includes tools for model development, data processing, training, inference, and evaluation. Our aim is to help agents perceive their environment, choose useful actions, and improve through feedback.

Explore the first release​

• Valen source repository
• Valen Preview 0923
• General 100k training data

For newer model releases, visit the model overview.

Source: Valen release notes.