# Notice

HexAnima MOSS-Image packs. These files are converted and quantised derivatives of the open models listed below. **Every component is licensed under the Apache License 2.0** (see `LICENSE`).
HexAnima is an independent, unofficial project. It is not affiliated with, endorsed by or sponsored by OpenMOSS, Unsloth, the Qwen team / Alibaba, Black Forest Labs, Qualcomm or the other authors named here.

## Components

| What | Files in this pack | Original | Author | License |
|---|---|---|---|---|
| MOSS-Image-SFT (5.1B text-to-image transformer) | `dit/blk00.bin` … `dit/blk41.bin`, `glue.bin` | https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT (GGUF quantisation by Unsloth: https://huggingface.co/unsloth/MOSS-Image-SFT-GGUF, file `MOSS-Image-SFT-UD-Q5_K_XL.gguf`) | OpenMOSS; quantisation Unsloth AI | Apache-2.0 |
| rapid-moss 4-step adapter (rank-128 LoRA, time path and sampler matrix) | LoRA weights inside `dit/blk*.bin`, `glue.bin` (`latent_mix.Q`, `step_codes`, time embedding, biases) | https://huggingface.co/darask0/MOSS-Image_4-step_adapter_lora (continues https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd) | darask0, lodestones | Apache-2.0 |
| Qwen3.5-9B (text encoder) | `te/te00.bin` … `te32.bin`, `embd.q4k`, `ref/*.bin` (text projection and refiner blocks come from MOSS-Image-SFT) | https://huggingface.co/Qwen/Qwen3.5-9B (GGUF quantisation by Unsloth: https://huggingface.co/unsloth/Qwen3.5-9B-GGUF, file `Qwen3.5-9B-UD-Q4_K_XL.gguf`) | Qwen team, Alibaba Cloud; quantisation Unsloth AI | Apache-2.0 |
| Qwen tokenizer | `tok_vocab.txt`, `tok_merges.txt`, `tok_special.txt` | `tokenizer.json` of https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT | Qwen team | Apache-2.0 |
| FLUX.2 small decoder (VAE decoder) and its batch-norm statistics | `vae/vae_<W>x<H>.bin`, `glue.bin` (`bn.mean`, `bn.var`) | https://huggingface.co/black-forest-labs/FLUX.2-small-decoder | Black Forest Labs | Apache-2.0 |

## Changes made to the original works

- The Unsloth GGUF files were dequantised and re-packed into 32x32 tiles for the Hexagon matrix unit: 8-bit blocks with one fp16 scale per column and 32 weights, or (Q4_K tensors of the text encoder only) a lossless 4-bit affine form. K-quant groups of 32 are kept; Q5_K / Q6_K / Q8_0 / IQ4_XS tensors are stored as 8-bit blocks.
- The rapid-moss LoRA is **not merged** into the weights (its delta is smaller than bf16 rounding); its two matrices are stored next to every projection and applied as extra input tiles.
- The FLUX.2 small decoder was converted to a Qualcomm HTP (V69, fp16) context binary for fixed image sizes.
- The weights are not usable with any other software.

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