Wan2.2-TI2V-5B — MLX 8-bit (q8)

MLX-format, 8-bit quantized conversion of Wan-AI/Wan2.2-TI2V-5B for native inference on Apple Silicon via mlx-video.

TI2V-5B is the single-model (5B) Text+Image-to-Video variant of Wan2.2, using the Wan2.2 VAE (z_dim=48). It runs comfortably at 720p (1280×704), 24fps on a 64 GB Apple Silicon machine, where the heavier 14B dual-model (I2V-A14B) struggles.

Contents

Total ≈ 18 GB.

Quantization

• 8-bit, group_size=64, applied to transformer Linear layers (self/cross-attention Q/K/V/O + FFN). Embeddings, norms, and the output head remain bf16.
• Converted with mlx_video.models.wan_2.convert --quantize --bits 8 --group-size 64.

Usage (mlx-video)

pip install git+https://github.com/Blaizzy/mlx-video.git
# or: uv pip install git+https://github.com/Blaizzy/mlx-video.git

huggingface-cli download <THIS_REPO_ID> --local-dir ./Wan2.2-TI2V-5B-MLX-Q8

python -m mlx_video.models.wan_2.generate \
  --model-dir ./Wan2.2-TI2V-5B-MLX-Q8 \
  --image ./start.png \
  --prompt "the subject waves hello, warm sunlight, film grain" \
  --width 1280 --height 704 --num-frames 81 \
  --steps 40 --guide-scale 5.0 \
  --output-path out.mp4

• Resolution must be divisible by 32. Frame count must be 4n+1. Output is 24 fps.
• --image is optional (TI2V also does pure text-to-video).

Hardware

Validated on an Apple Silicon Mac (64 GB unified memory): 720p × 41 frames × 20 steps ≈ 15 min, stable memory. 32 GB+ recommended.

License & Attribution

This is a derivative of Wan-AI/Wan2.2-TI2V-5B, released under Apache-2.0. The Apache-2.0 license and NOTICE are included. All credit for the base model goes to the Wan-AI / Alibaba team. MLX conversion via Prince Canuma's mlx-video. This repository only quantizes and re-packages the weights for MLX; the model architecture and training are unchanged.