PocketAiHub/PocketAI-Video-MLX-Wan2.2-A14B

Wan2.2 T2V A14B Q4

Apple-Silicon-native Q4 MLX packaging of Wan2.2 T2V A14B with dual high/low-noise experts, a bundled offline tokenizer, a shared T5/VAE, and a tested command-line runtime.

This is a community derivative assembled by PocketAI. It is not an official Wan-AI or SceneWorks release, and no endorsement is implied. PocketAI did not train the base model or perform the original SceneWorks conversion.

Available precision variants

Precision	Revision
Q4	main
Q8	q8
BF16	bf16

Provenance

• Base model: Wan-AI/Wan2.2-T2V-A14B
• MLX conversion: SceneWorks/wan2.2-t2v-a14b-mlx
• Runtime: mlx-video
• PocketAI changes: MODIFICATIONS.md

Included files

Folder	Contents	Unique size
common/	UMT5-XXL encoder, tokenizer, Wan z16 VAE	11.07 GiB
q4/	Q4 high/low-noise experts, group size 64	15.61 GiB

Model data totals 26.68 GiB. The bundled runtime resolves the shared support models from common/ automatically.

Requirements

• Apple Silicon Mac
• Python 3.11, 3.12, or 3.13
• ffmpeg available on PATH
• Approximately 26.68 GiB disk space plus the Python environment and generated videos
• 24 GB unified memory for short low-guidance tests; 128 GB was used for the validated 81-frame CFG benchmark

Install

hf auth login
hf download PocketAiHub/PocketAI-Video-MLX-Wan2.2-A14B --revision main \
  --local-dir pocketai-wan22-mlx-q4
cd pocketai-wan22-mlx-q4
python3 -m venv .venv
.venv/bin/pip install .

Install ffmpeg if needed:

brew install ffmpeg

Generate a five-second video

.venv/bin/pocketai-wan-video \
  --model-dir q4 \
  --prompt "A chestnut horse runs across a sunlit field, side tracking camera, photorealistic." \
  --width 512 --height 320 --num-frames 81 \
  --steps 10 --guide-scale 3,4 --scheduler unipc \
  --seed 42 --output-path horse.mp4

Wan uses 16 fps in this package. 81 frames encode to 5.0625 seconds. Frame counts must follow 4n+1, and width/height should be divisible by 32.

Quick smoke test

.venv/bin/pocketai-wan-video \
  --model-dir q4 \
  --prompt "A paper boat moves across calm water." \
  --width 256 --height 256 --num-frames 5 \
  --steps 1 --guide-scale 1 --scheduler euler \
  --no-compile --seed 2301 --output-path smoke.mp4

PocketAI benchmark

Apple M5 Max, 128 GB unified memory; identical prompt and seed; 512×320, 81 frames, 16 fps, 10 UniPC steps, guidance 3/4:

Precision	End-to-end	Denoising	Max process RSS	Observed result
Q4	10:25	9:06	16.37 GiB	Fastest; good prompt adherence
Q8	11:17	10:07	29.46 GiB	Sharpest; best quality-oriented option
BF16	11:03	10:15	41.48 GiB	Q8-like speed; no quality win in this seed

Q4 is the recommended default. Q8 is the higher-detail option. BF16 is included for experimentation and maximum numerical precision, not because one test establishes it as perceptually better.

Watch the synchronized comparison. Machine-readable measurements are in benchmark/metrics.json.

Limitations

• Wan2.2 A14B is large and slow. A separate 768×512 Q4 reference took 36:53 end to end on the test machine.
• Same-seed outputs can diverge across precisions; divergence is not itself a quality ranking.
• This release has not been validated on Intel Macs, iOS, or machines with less memory.
• Ten UniPC steps are a speed-oriented setting. More steps can improve results at substantial additional cost.
• Review the Apache-2.0 license and applicable laws before use.

License

The model derivative is distributed under the Apache License 2.0 reproduced in LICENSE. The bundled mlx-video runtime is MIT-licensed; its license is reproduced in LICENSE-RUNTIME-MIT. See THIRD_PARTY_NOTICES.md for attribution.