README.md

---
license: other
license_name: kolors-model-license
license_link: LICENSE
library_name: mlx
pipeline_tag: text-to-image
base_model: Kwai-Kolors/Kolors-diffusers
tags:
  - mlx
  - kolors
  - text-to-image
  - quantized
---

# Kolors — MLX quantization tiers (SceneWorks)

Pre-quantized MLX turnkey tiers of [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers)
for the SceneWorks native (`mlx-gen`) runtime. Each tier subdir is a self-contained snapshot — SDXL-style
U-Net + **ChatGLM3-6B** text encoder + SDXL VAE + baked fast tokenizer — that loads **packed**, with no
dense transient and no install-time conversion peak.

| Tier | Subdir | Size | Notes |
|------|--------|------|-------|
| Q4 (default) | `q4/`  | ~6.0 GiB | U-Net + ChatGLM3 packed 4-bit; dense VAE |
| Q8           | `q8/`  | ~9.7 GiB | U-Net + ChatGLM3 packed 8-bit; dense VAE |
| bf16         | `bf16/`| ~17 GiB  | dense 16-bit weights |

**Quantization:** group-wise affine, group size 64, weight-only, byte-identical to the runtime's
load-time seam. The VAE is never quantized (it runs f32). The fast `tokenizer.json` is derived from the
ChatGLM3 slow tokenizer ([`SceneWorks/kolors-chatglm3-tokenizer`](https://huggingface.co/SceneWorks/kolors-chatglm3-tokenizer))
and baked into each tier.

## License & attribution

Model weights are Kolors, © Kwai-Kolors, under the **Kolors Model License Agreement** — see
[`LICENSE`](LICENSE) (verbatim). Commercial use requires a separate license from the original licensor.
Re-hosted by SceneWorks for runtime provisioning; the license and its use-restrictions travel with the
weights and bind downstream recipients (per §3(a) of the license).