hexanima-qwen-packs / NOTICE.md
ayanoby's picture
add NOTICE.md
659aaaa verified
|
Raw History Blame Contribute Delete
3.13 kB

Notice

HexAnima Qwen-Image pack. These files are converted and quantised derivatives of the models listed below. HexAnima is an independent, unofficial project; it is not affiliated with, endorsed by or sponsored by the Qwen team / Alibaba, LightX2V, Unsloth, Qualcomm or the other authors named here. The transformer, the Lightning adapter and the VAE are licensed under the Qwen RESEARCH LICENSE AGREEMENT (see LICENSE): non-commercial use only. The text encoder components are under the Apache License 2.0.

Built with Qwen.

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

Components

What Files in this pack Original Author License
Qwen-Image-2.1-Turbo transformer (20B) dit/blk00.bin … blk31.bin, glue.bin https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo (GGUF quantisation by Unsloth, file qwen-image-2.1-turbo-Q5_K_S.gguf) Qwen team, Alibaba; quantisation Unsloth AI Qwen Research License
Lightning v2-RL rank-64 adapter LoRA rows inside dit/blk*.bin, glue.bin https://huggingface.co/lightx2v/Qwen-Image-2.1-Lightning (v2 RL, rank 64) LightX2V No license is declared in that repository; as an adapter of Qwen-Image-2.1 it is treated as a derivative of the Qwen Materials, i.e. Qwen Research License (non-commercial)
Qwen3-VL-8B-Instruct (text encoder) te/te00.bin … te35.bin, embd.q4k https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct (GGUF quantisation by Unsloth, UD-Q4_K_XL) Qwen team; quantisation Unsloth AI Apache-2.0
Qwen3-VL vision tower vit/vit00.bin … vit26.bin, vit/vit_glue.bin same model (mmproj-F16.gguf / visual weights) Qwen team Apache-2.0
Qwen-Image-2.1 VAE vae/vae_t32.bin, vae/enc_t32.bin https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo (VAE) Qwen team Qwen Research License
Tokenizer tok_*.txt tokenizer.json of Qwen3-VL-8B-Instruct / Qwen-Image-2.1 Qwen team Apache-2.0

Changes made to the original works (Qwen Research License, section 3b)

  • The transformer GGUF was dequantised and re-packed into 32x32 tiles for the Hexagon matrix unit: Q4_K tensors in a lossless 4-bit affine form (including the 4-bit LoRA rows), Q5_K / Q6_K / Q8_0 tensors as 8-bit tiles with one fp16 scale per column and 32 weights. The weights are not otherwise changed.
  • The Lightning LoRA is not merged into the weights (its delta is about 0.1-0.2 % of them); its matrices are stored next to every projection and applied as extra tiles.
  • The text encoder and vision tower were re-packed into 4-bit and 8-bit tiles; the vision tower's 72-wide attention heads are zero-padded to 128.
  • The VAE encoder and decoder were re-written as 4-D graphs and compiled to fp16 Hexagon (HTP V69) contexts for fixed tile sizes.
  • Tokenizer files were converted to plain text lists.
  • The weights are not usable with any other software.

No other changes were made. The licenses and copyright notices of the original works apply to these derivatives; see LICENSE.