zing-0.5

<img src="https://huggingface.co/seedleap/zing-0.5/resolve/main/assets/zing-logo.svg" alt="Zing" height="52" style="vertical-align: middle;"> <span style="font-size: 1.2em;">Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control</span>

Project Page · GitHub · Tech report

Zing-0.5 is a 5B causal world model developed by the Seedleap.ai team (涌跃智能) for real-time interaction. It continuously predicts the visual world from its current state while text prompts and keyboard actions alter the scene, motion, and future evolution during generation.

Zing-0.5 supports text-initialized generation, single-image initialization, prompt changes during a rollout, and continuous W/A/S/D/I/J/K/L keyboard control. Causal KV caching and four-step DMD sampling enable responsive long-horizon generation on a single GPU.

Model Files

Download the model repository with the following directory structure:

Zing-0.5/
├── generator/
│   └── model.pt
└── pretrained/
    ├── text_encoder/
    ├── tokenizer/
    └── vae/

generator/model.pt must directly contain the generator state dict. Parameter names and shapes are loaded strictly.

Inference

Use the standalone inference code from the Zing GitHub repository:

git clone https://github.com/seedleap/zing-world-model.git
cd zing-world-model

ZING_MODEL=/path/to/Zing-0.5
CUDA_VISIBLE_DEVICES=0 \
ZING_PYTHON=/path/to/python \
bash run.sh \
  --pretrained-dir "$ZING_MODEL/pretrained" \
  --checkpoint "$ZING_MODEL/generator/model.pt" \
  --messages examples/case3_action_t2v.jsonl \
  --output-dir outputs/case3 \
  --seed 0

The code repository includes ready-to-run Action T2V and Action TI2V JSONL examples with the required reference images.

GPU Memory

The default 97/9 configuration has been validated on a single NVIDIA H100 80 GB. Use 33/5 on GPUs with less memory. Full-history attention is available with --local-attn-size -1 --sink-size 0.

Output

Each JSONL row produces one result named from sample_id. Multi-frame rollouts are saved as H.264 MP4 at 24 FPS.

Limitations

Long rollouts may exhibit visual drift or physical inconsistencies. Action responsiveness can vary with scene content and viewpoint.

License

Zing-0.5 is released under the Apache License 2.0.