| --- |
| license: apache-2.0 |
| base_model: Wan-AI/Wan2.2-I2V-A14B |
| library_name: mlx |
| pipeline_tag: image-to-video |
| tags: |
| - mlx |
| - apple-silicon |
| - wan2.2 |
| - image-to-video |
| - video-generation |
| - mixture-of-experts |
| --- |
| |
| # Wan2.2-I2V-A14B — MLX (bf16) |
|
|
| Native **MLX** (Apple Silicon) conversion of [Wan-AI/Wan2.2-I2V-A14B](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B), packaged as a turnkey, self-contained snapshot for the [SceneWorks](https://huggingface.co/SceneWorks) app. Downloading this repo replaces the previous "download the native checkpoint and convert on-device" first-run step. |
|
|
| Wan2.2 A14B is a **high/low-noise mixture-of-experts** image-to-video model — two transformers switched at the noise boundary. |
|
|
| ## Contents (self-contained, bf16) |
| | file | what | |
| |---|---| |
| | `high_noise_model.safetensors` | high-noise expert DiT (~28.6 GB) | |
| | `low_noise_model.safetensors` | low-noise expert DiT (~28.6 GB) | |
| | `t5_encoder.safetensors` | UMT5-XXL text encoder (~11.4 GB) | |
| | `vae.safetensors` | Wan z16 VAE | |
| | `tokenizer.json` | UMT5 tokenizer | |
| | `config.json` | architecture config (`dual_model: true`, `in_dim: 36` image-concat) | |
|
|
| Quantization (Q4/Q8) is applied **at load** by the engine — these weights are full bf16. |
|
|
| ## Provenance |
| - **Source:** `Wan-AI/Wan2.2-I2V-A14B` (Apache-2.0). |
| - **Converted with:** the SceneWorks native Rust MLX converter (`mlx-gen-wan`, converter id `wan_i2v_14b`), dtype `bfloat16`, dense (no baked-in quant). |
| - Lean snapshot — only the files the MLX engine loads. |
|
|
| ## License |
| **Apache-2.0**, inherited from the upstream model. This repository redistributes a converted copy of the upstream Apache-2.0 weights, with attribution, as permitted by that license. See the [source model card](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B). |
|
|
