User repository on GitHub
Train a Jev-like multimodal model by yourself. System One Model, now with vision.
- Forks
- 82
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
Valen
A multimodal decision model that scores text, image, and video candidates with a shared head on Qwen3.5, plus SFT and experimental RLCD training code.
- Stars
- 861
- Language
- Jupyter Notebook
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- Liuziyu77
- Last verified
- Sep 2026
About this project
A multimodal decision model inspired by Jev — text, images and video in; decision probabilities out. English · 简体中文 Intro · News · Demos · Models & datasets · Eval results · Quick start.
Quoted from the project's README. Read the full README.
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- Backbones are Qwen3.5-0.8B or 2B. A shared decision head scores candidates without generating answer tokens.
- The repo includes model code, data processing, SFT, experimental RLCD, inference, and evaluation.
- Valen-Preview-0923 solved a Sokoban level in 9 decisions and 1.13 seconds of cumulative latency.
- On the same level, Qwen3.8-27B-FP8 took 198.05 seconds in thinking mode and failed with thinking off.
- Decision confidence on a blur demo fell from 91.6% on the clear image to 19.2% at the strongest blur.
Quickstart
bash
hf download Valen-Team/Valen-Preview-0923 --local-dir models/Valen-Preview-0923
python -m valen.inference --checkpoint models/Valen-Preview-0923 \
--data data/smoke/train.jsonl --output output/sft_warmup/predictions.jsonlWatch out
Apache-2.0 for the code. Base models and datasets keep their own licenses. Inference needs both the Valen checkpoint and the Qwen3.5-2B base. Not a TypeSafe client.
Questions about this entry
Short answers drawn from the linked project or page.
A shared decision head over a Qwen3.5-0.8B or 2B backbone. It scores the candidates you supply without generating answer tokens, so there is no decoding step and no prose to parse. The repo also carries data processing, SFT, experimental RLCD training, and inference and evaluation commands.
Only declared options. The labeled record format is a state plus questions, each with a criteria map, and targets holding one probability per candidate. The README's example is a choice question; Score and Noul are not demonstrated in the excerpt.
They are one level. Valen-Preview-0923 solved it in 9 decisions and 1.13 s of cumulative latency; Qwen3.8-27B-FP8 took 198.05 s in thinking mode and failed with thinking off. Broader accuracy and latency numbers for General (5,000 questions) and Sokoban (500 single-step questions from 100 levels) sit in a figure, not in text you can check here.
The quick start asks for both downloads and states that running Valen requires the checkpoint and the Qwen3.5-2B base model, because the shared head needs the backbone weights separately from the Valen weights. The README does not explain the reason any further. data/smoke holds small synthetic questions for checking the pipeline, and the code is Apache-2.0 while base models and datasets keep their own licenses.
Three sizes: 0.8B, 2B and 4B, all Qwen3.5 plus a shared Mixer head. On VisualDecisionBench-Image accuracy runs 75.94, 80.41 and 83.06 percent; on Video 79.03, 82.85 and 87.36; and on public JevBench 77.92, 83.55 and 83.12, where a published InternDecision 4B scores 86.58.
The repository includes data processing, SFT and experimental RLCD training, inference and evaluation commands, and support for training on your own data, with Flash Attention 2 and multi-node multi-GPU training added in September 2026. The code is Apache-2.0, but the base models and datasets keep their own licenses.
There is a Hugging Face Space, Valen-Preview-0923, to try it without setup. Running it yourself needs the Valen checkpoint plus the separate Qwen3.5 base, a CUDA device, trust_remote_code, and a call to model.predict with execution shared_state or question; image and video inputs follow the documented data format.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsX: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
More Jev resources like this
An open-model semantic-if engine positioned as a local Jev-style decision stack, separate from the original SemIf repo but sharing the same author.
Jev Models & Reimplementations#community#python#open-models
Community
TheoLeeCJ/SemIf-OpenJevPython
TheoLeeCJ/SemIf-OpenJevPythonA local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonA Python server in front of llama-server that reads one-token label probabilities and exposes them as POST /v1/systemone.
Jev Models & Reimplementations#community#python#open-models
Community
khimaros/verdictPython
khimaros/verdictPythonMapika's vision decider: decider-2b v5 text weights transplanted into the Qwen3.5-2B vision-language model, tuned so an image plus a lettered question returns a calibrated probability at one answer slot with no generation. Breakout from pixels scores 41, above its teacher's 22.
Jev Models & Reimplementations#community#open-models#local
Community
Ruiruiz30's Apple Silicon MLX conversion of Jev-Omni: 4-bit affine language weights on a Mac mini M4 16GB, about 7 GB peak Metal memory and roughly 963 ms text decisions, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.