Wan2.2-I2V-A14B — MLX 8-bit (q8)

MLX-format, 8-bit quantized conversion of Wan-AI/Wan2.2-I2V-A14B for native inference on Apple Silicon via mlx-video.

I2V-A14B is the 14B dual-model (MoE) Image-to-Video variant of Wan2.2: a high-noise expert and a low-noise expert are swapped across the denoising schedule.

Note on hardware: This is a large model. The two 14B experts (~28 GB q8) plus the 11 GB T5 encoder make 720p generation very heavy on a 64 GB machine (heavy swapping). For 64 GB Apple Silicon, the lighter [Wan2.2-TI2V-5B MLX q8] is far more practical at 720p. This A14B build targets machines with more headroom (96–128 GB+).

Contents

Total ≈ 40 GB.

Quantization

• 8-bit, group_size=64, applied to transformer Linear layers (self/cross-attention Q/K/V/O + FFN) in both experts. 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

huggingface-cli download <THIS_REPO_ID> --local-dir ./Wan2.2-I2V-A14B-MLX-Q8

python -m mlx_video.models.wan_2.generate \
  --model-dir ./Wan2.2-I2V-A14B-MLX-Q8 \
  --image ./start.png \
  --prompt "the person slowly turns and smiles, cinematic, natural lighting" \
  --width 1280 --height 704 --num-frames 81 \
  --steps 40 --guide-scale "3.5,3.5" \
  --tiling aggressive \
  --output-path out.mp4

• Resolution must be divisible by 16. Frame count must be 4n+1. Output is 16 fps.
• Dual-model → guidance is a high,low pair (--guide-scale "3.5,3.5").
• Use --tiling aggressive (and lower resolution) on memory-constrained machines.

License & Attribution

This is a derivative of Wan-AI/Wan2.2-I2V-A14B, 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.