| --- |
| 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). |
|
|
