Organization repository on GitHub
VideoJev: System One Model for Video Understanding and Interaction
- Forks
- 7
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
VideoJev
A System One model family for video: VideoJev-4B and VideoJev-35B-A3B perform temporal grounding and multiple-choice video QA by scoring candidates directly, returning probabilities instead of generated text.
- Stars
- 162
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- mcg-nju
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open source
Highlights
- Grounding returns per-interval sigmoid probabilities with a default threshold of 0.25; choice returns softmax probabilities and a zero-based answer_index.
- Two checkpoints on Hugging Face: VideoJev-4B on Qwen3.5-4B and VideoJev-35B-A3B on Qwen3.6-35B-A3B, both BF16 safetensors.
- The 10-second quick start loads ZhuYuhan/VideoJev-Demo eating_spaghetti.mp4 and calls predict_video at fps=2.0, max_frames=2048, context=8192.
- Training is two stages, head warmup then joint, with the vision tower frozen; each stage runs one epoch with a global batch of 64 packs.
- Evaluation runs through scripts/evaluate_hf.py; QA uses a fixed 256K context and the same command works for both checkpoints.
Quickstart
bash
pip install -r requirements.txt
source scripts/env.sh
python scripts/evaluate_hf.py \
--checkpoint ZhuYuhan/VideoJev-4B \
--manifest /path/to/eval.jsonl --out runs/eval4bWatch out
Repo and README reviewed 2026-10-15. The README presents results as paper tables (images) and model cards, so this entry quotes no accuracy numbers, and requirements are steep: Python 3.12, PyTorch 2.8 with CUDA 12.8, FlashAttention 2, causal-conv1d, and for 35B training a compiled grouped-GEMM extension through XTuner.
Questions about this entry
Short answers drawn from the linked project or page.
A System One style model family for video understanding: it performs temporal grounding and multiple-choice video QA by scoring candidate answers directly, without generating text. The repo ships training, inference, and evaluation code around two checkpoints.
It is an independent System One model in the same non-generative style as Jev, from the MCG-NJU group: grounding and QA answers come back as probabilities rather than text. The repository does not mention TypeSafe, and no affiliation is claimed.
VideoJev-4B built on Qwen/Qwen3.5-4B and VideoJev-35B-A3B built on Qwen/Qwen3.6-35B-A3B, both BF16 safetensors hosted as ZhuYuhan/VideoJev-4B and ZhuYuhan/VideoJev-35B-A3B. The repository is Apache-2.0.
Python 3.12, PyTorch 2.8 with CUDA 12.8, Triton 3.4, FlashAttention 2, and causal-conv1d. Training adds XTuner, and the 35B path also needs its compiled grouped-GEMM and token-permutation extension, with torchrun examples using eight GPUs.
The README quick start downloads a demo video from the ZhuYuhan/VideoJev-Demo dataset repo, loads the processor and model with trust_remote_code, calls prepare_for_inference, then predict_video with a questions list and prints the JSON results.
Grounding results give per-candidate sigmoid probabilities and an interval in seconds, thresholded at 0.25 by default and overridable through predict_video. Choice results give softmax probabilities over the options plus a zero-based answer_index and the chosen text.
Yes in principle: prepare JSONL data with the documented tooling, generate configs for either model, then run the two torchrun stages, head warmup and joint. Vision stays frozen in both stages, and the 4B export produces a Hugging Face format checkpoint under joint/hf_checkpoint/.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
More Jev resources like this
An independent fork of Laya that replaces its text encoder with a small vision-language model, giving calibrated typed decisions over an image plus text in one forward pass from a 201M-parameter SmolVLM-256M backbone.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
r33drichards/laya-visionPython
r33drichards/laya-visionPythonA 144.3M-parameter bilingual typed decision model fine-tuned from jhu-clsp/mmBERT-small: it answers noul, choice, and score questions in one forward pass with calibrated confidence and zero output tokens, and it runs on CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
An open, local attempt to bring Jev-style typed decisions to images: a Qwen3-VL backbone returns numerical distributions for independent choice, score and noul questions without generating answer text.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
andrueandersoncs/visual-jevPython
andrueandersoncs/visual-jevPythonA Japanese System One decision model built on modernbert-ja-310m: send text and typed questions and get typed answers with probabilities from one forward pass, no text generation, at 314.6M parameters that also run on a CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
hiroki-abe-58/sokudanPython
hiroki-abe-58/sokudanPythonOpen, CPU-sized Jev-style decision model: give it a situation, a question and options, and Decima returns calibrated probabilities in roughly 20-50 ms per decision as int8 ONNX, with no GPU, no API key and nothing leaving the machine.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
amyrmahdy/decimaPython
amyrmahdy/decimaPythonHAL-X AI's 322M non-autoregressive decision model returns calibrated answers in one forward pass of about 10 ms, with native number, excerpt and citation outputs beyond Jev's choice, noul and score types. Apache-2.0, post-trained in 18 languages.
Jev Models & Reimplementations#community#catalog-import#madewithjev
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.