Notice
HexAnima Mage-Flow-Edit-Turbo pack. These files are converted and quantised derivatives of the model listed below. HexAnima is an independent, unofficial project; it is not affiliated with, endorsed by or sponsored by Microsoft, the Qwen team / Alibaba or Qualcomm.
Components
| What | Files in this pack | Original | Author | License |
|---|---|---|---|---|
| Mage-Flow-Edit-Turbo transformer (12 double-stream blocks, 4-step distilled) | dit/blk00.bin … blk11.bin, glue.bin |
https://huggingface.co/microsoft/Mage-Flow-Edit-Turbo | Microsoft | MIT |
| Mage-VAE (encoder and decoder) | vae.bin, vaepack.bin, vaepack.idx |
same repository (vae) |
Microsoft | MIT |
| Qwen3-VL-4B text encoder (language model) | te/te00.bin … te35.bin, te/embd.q8, te/tefinal.bin |
same repository (text_encoder, a Qwen3-VL-4B checkpoint) |
Qwen team, Alibaba (as shipped by Microsoft) | Apache-2.0 |
| Qwen3-VL-4B vision tower | vit/vit00.bin … vit23.bin, vit/vit_glue.bin |
same repository (text_encoder) |
Qwen team | Apache-2.0 |
| Tokenizer | tok_*.txt |
tokenizer.json of the text encoder |
Qwen team | Apache-2.0 |
Changes made to the original works
- The transformer weights were re-packed into 32x32 tiles for the Hexagon matrix unit with 8-bit values and one fp16 scale per column and 32 weights. The adaLN modulation of the fixed 4-step schedule was computed in advance and stored in
glue.bin, so the modulation layers are not part of the pack. - The text encoder and the vision tower were re-packed the same way (8-bit tiles).
- The VAE weights were re-ordered for the app's own decoder / encoder kernels (fp16).
- Tokenizer files were converted to plain text lists.
- The weights are not usable with any other software.
No other changes were made. The official pipeline's content filter and watermark are not part of the app's runner.