Latent-VC-9B

Related Resources

Model Details

• Architecture: Qwen3_5ForConditionalGeneration (Qwen3.5-VL backbone)
• Hidden size: 4096  |  Layers: 32 (hybrid linear + full attention)  |  Vision depth: 27
• Dtype: bfloat16
• Parameters: ~9B
• Special capability: Latent Visual Compression (LVC) tokens for latent video reasoning (lvc_temperature=0.07, loss_lvc_fct=cosine)
• Training: SFT (Stage 1) → GRPO (Stage 2); this checkpoint corresponds to the GRPO checkpoint-800.

Files

This model uses custom LVC components. For training and LVC reasoning inference, please use the code from the Latent-VC repository.

Usage

Download

bash
Copy code
pip install -U "huggingface_hub[cli]"
hf download BRZ911/Latent-VC-9B --local-dir ./Latent-VC-9B

Evaluation

The included scripts evaluate the model on six video benchmarks (VideoMME, MVBench, TempCompass, VideoMMMU, VSIBench, MMVU). The evaluation data is hosted in the companion dataset BRZ911/Latent-VC-Data under Eval/ — download and extract it so that an Evaluation/ directory (with eval_<dataset>.json and the per-benchmark videos) is available, then run:

bash
Copy code
# MODEL_PATH : path to the downloaded weights (this repo)
# EVAL_DIR   : path to the extracted Evaluation/ directory (default: ./Evaluation)
# DATASETS   : comma-separated subset of
#              videomme,mvbench,tempcompass,videommmu,vsibench,mmvu
MODEL_PATH=./Latent-VC-9B \
EVAL_DIR=./Evaluation \
DATASETS=mmvu \
bash eval_all_benchmarks.sh

See the Latent-VC repository for the full inference/training environment and dependencies.

Benchmarks

The model is evaluated on the following benchmarks (data in BRZ911/Latent-VC-Data):

Citation

If you find this model or dataset useful, please cite the project:

bibtex
Copy code
@misc{latentvc,
  title  = {Latent-VC: Latent Visual Compression for Efficient Video Reasoning},
  author = {BRZ911},
  year   = {2025},
  url    = {https://github.com/BRZ911/Latent-VC}
}