Yume-1.5-5B-720P-MLX-4bit

Yume 1.5 5B 720P — MLX (4-bit Quantized)

4-bit quantized MLX weights for Yume 1.5, a 5B-parameter interactive world generation model. Runs natively on Apple Silicon via mlx-video.

This is the quantized variant — the transformer is 2.7GB instead of 9.7GB, making it more feasible on machines with limited memory. For full-precision BF16 weights, see Yume-1.5-5B-720P-MLX.

Yume 1.5 is fine-tuned from Wan2.2 TI2V-5B and supports both text-to-video and image-to-video generation.

Quick Start

pip install mlx-video

# Image-to-Video
python -m mlx_video.models.wan_2.generate \
    --model-dir ckurasek/Yume-1.5-5B-720P-MLX-4bit \
    --image my_photo.jpg \
    --prompt "A dog runs through the scene" \
    --width 832 --height 480 --num-frames 81 \
    --steps 30 --seed 42 \
    --output-path output.mp4

# Text-to-Video (omit --image)
python -m mlx_video.models.wan_2.generate \
    --model-dir ckurasek/Yume-1.5-5B-720P-MLX-4bit \
    --prompt "A cinematic shot of a mountain landscape at sunset" \
    --width 832 --height 480 --num-frames 81 \
    --steps 30 --seed 42 \
    --output-path output.mp4

The model directory is auto-downloaded from Hugging Face on first use.

Quantization Details

• Method: 4-bit affine quantization (group_size=64)
• Quantized layers: Self-attention Q/K/V/O, cross-attention Q/K/V/O, FFN fc1/fc2 (300 layers)
• Preserved in BF16: Embeddings, norms, output head, modulation
• T5 encoder and VAE: Unquantized (BF16 and float32 respectively)

Files

Model Details

• Architecture: Wan2.2 TI2V-5B (DiT backbone, T5 text encoder, 3D VAE)
• Parameters: 5B (transformer) + 5.7B (T5) + 72M (VAE)
• Resolution: 720P native (1280x704), works at lower resolutions
• FPS: 24

Attribution

• Yume 1.5: Paper | GitHub | Project Page
• Wan2.2: GitHub
• mlx-video: GitHub

Citation

@article{mao2025yume15,
  title={Yume-1.5: A Text-Controlled Interactive World Generation Model},
  author={Mao, Xiaofeng and Li, Zhen and Li, Chuanhao and Xu, Xiaojie and Ying, Kaining and He, Tong and Pang, Jiangmiao and Qiao, Yu and Zhang, Kaipeng},
  journal={arXiv preprint arXiv:2512.22096},
  year={2025}
}

License

Apache 2.0 (same as the original Yume model).