Wan2.2-I2V-A14B — MLX bf16 (relay-shedding)
Self-contained MLX bf16 conversion of Wan-AI/Wan2.2-I2V-A14B, the Quality tier in Videoboom. Both denoise experts are kept unquantized so mlx-video's relay-shedding can load one expert at a time — measured ~32.6 GB peak (fits 48 GB unified), vs the ~67.7 GB of Q4/Q8 that keep both experts resident. Pairs with the Wan2.2-Lightning 4-step I2V LoRA for a 4-step fast path.
| file | size |
|---|---|
high_noise_model.safetensors |
28.6 GB |
low_noise_model.safetensors |
28.6 GB |
t5_encoder.safetensors |
11.4 GB |
vae.safetensors |
0.5 GB |
License Apache-2.0 (inherited from Wan-AI/Wan2.2-I2V-A14B).
- Downloads last month
- 283
Hardware compatibility
Log In to add your hardware
Quantized
Model tree for lBroth/Wan2.2-I2V-A14B-MLX-bf16
Base model
Wan-AI/Wan2.2-I2V-A14B