---
license: other
tags: [hexanima, qualcomm, hexagon, npu, anima]
---
# HexAnima packs for Snapdragon 8 Gen 2 (Hexagon V73)

Shared files that [HexAnima](https://github.com/ayanoby/hexanima) downloads on a Snapdragon 8 Gen 2 phone (`common_v73/`):
the Qwen3-0.6B text encoder of Anima as four QNN HTP context binaries compiled for Hexagon V73 (512 token slots),
the TAEHV (`taew2_1`) preview decoder as a V73 context, the Qwen3 token embedding table (fp16) and the text-encoder header stub.
The model weights themselves (MiaoMiao packs) are downloaded from `ayanoby/anima-npu-packs`.

Licenses: Qwen3-0.6B text encoder and embeddings: Apache 2.0 (Qwen team), as shipped with Anima (CircleStone Labs Non-Commercial License).
TAEHV: see github.com/madebyollin/taehv. Non-commercial use only, like the Anima models.
Built with `tools/build_ctx.sh` / `tools/make_common_pack.sh` of the HexAnima V73 port.

## adetailer/face_yolov8n.safetensors

Face detector for HexAnima's ADetailer option: `face_yolov8n.pt` from [Bingsu/adetailer](https://huggingface.co/Bingsu/adetailer),
converted with stable-diffusion.cpp's `scripts/convert_yolov8_to_safetensors.py` (BatchNorm fused into the convolutions).
The detector was trained with Ultralytics YOLOv8 (AGPL-3.0); see the original repository for its terms.