Valen
Multimodal System One: text, images, video in; Qwen3.5 plus shared decision head out. HF Valen-Preview-0923, Spaces demo, SFT and RLCD.
Overview
Valen (github.com/Liuziyu77/Valen, Apache-2.0) trains multimodal System One models: text, images, and video in; choice, score, and noul style probabilities out without answer-token decoding. Qwen3.5-0.8B or 2B backbones share a decision head; Valen-Preview-0923 ships on Hugging Face Valen-Team/Valen-Preview-0923 with Spaces demo, General 100k training data, and game eval sets. README Sokoban comparison claims Valen solves a level in 9 decisions with 1.13 s cumulative latency vs Qwen3.8-27B-FP8 at 198.05 s in thinking mode on the same puzzle (project demo, not independent lab proof). Blur-confidence GIF documents 91.6% confidence clear vs 19.2% at strongest blur. Repo includes SFT and experimental RLCD configs, valen.train, valen.inference, and valen.evaluate CLIs.
Problem: Text-only Von, Kev, and tev1 checkpoints cannot score visual state, yet multimodal agents still need thresholdable probabilities instead of long reasoning traces.
Built for: Researchers and builders training or serving Qwen3.5-backed System One models with image and video inputs via Hugging Face Valen-Preview-0923.
First indexed on Jev Directory: 2026-09-25
Creator and team
- Name
- Valen Team
- Handle
- @Liuziyu77
How Jev is used
- Role in the product flow
- Multimodal forward pass scoring supplied candidates with a shared decision head
- Primitives
- ChoiceScoreNoul
- State in
- JSONL records with messages content arrays for text, image_url, or video plus questions maps per README smoke and eval schemas.
- Decision out
- Probability distributions over listed criteria keys; zero generated answer tokens on the decision path.
- hf download Valen-Preview-0923 and matching Qwen3.5-2B base weights
- python -m valen.inference --checkpoint models/Valen-Preview-0923 on JSONL tasks
- python -m valen.evaluate for General 5k or game sets documented in README
- Optional SFT or RLCD training via configs/train per docs/technical.md
Valen extends the open decision model lane into vision: same candidate scoring story as jaredpalmer-kev and togethercomputer-tev1, but tensors include frames. tianyucodings-nanojev benchmarks tiny parallel heads on games; Valen targets general multimodal VQA plus Sokoban eval sets with published datasets. wfzyx-von remains text encoder speed king; Valen trades modality breadth for training complexity. Use the HF Space before you commit GPUs to a custom SFT run.
Sourced performance claims
- README Sokoban demo: Valen-Preview-0923 9 decisions and 1.13 s cumulative latency vs Qwen3.8-27B-FP8 198.05 s thinking on the same level (project-reported).Source: github.com/Liuziyu77/Valen README demo comparison
- Blur demo reports 91.6% decision confidence on clear image vs 19.2% at strongest Gaussian blur in project GIF.Source: github.com/Liuziyu77/Valen README blur-confidence section
- Public GitHub repo Liuziyu77/Valen had about 164 stars when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Valen-Preview-0923 weights and Qwen3.5 base pairing documented on HF
- Hugging Face Space demo and General 100k plus game dataset cards
- python -m valen.train, inference, and evaluate entry points
- SFT warmup and experimental RLCD configs under configs/train
- Project demos for Sokoban latency, blur confidence, and four parallel games
Stack
- Python 3.10+
- Qwen3.5-0.8B and 2B
- Hugging Face weights
- SFT and RLCD training
Pricing: Apache-2.0 code; inference and training spend your own GPU hours and HF bandwidth.
Links
FAQ
- Which base model do I need for Preview 0923?
- README quick start downloads Valen-Preview-0923 and Qwen/Qwen3.5-2B into local models paths before inference.
- Is Valen a hosted TypeSafe product?
- No. It is an open training and inference stack inspired by Jev with HF weights you serve yourself.
- How is this different from text-only Kev?
- Valen accepts image and video content in state messages and scores candidates with a shared multimodal decision head.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.