---
language:
- en
base_model:
- krea/Krea-2-Raw
base_model_relation: quantized
pipeline_tag: text-to-image
library_name: mlx
license: other
license_name: krea-2-community-license
license_link: https://huggingface.co/SceneWorks/krea-2-raw-mlx/blob/main/LICENSE.pdf
tags:
- mlx
- apple-silicon
- text-to-image
- diffusion
- krea-2
- quantized
---

# Krea 2 Raw — MLX (turnkey: bf16 / Q8 / Q4)

On-device, Apple-MLX-ready repack of **[krea/Krea-2-Raw](https://huggingface.co/krea/Krea-2-Raw)**, the
**undistilled** 12B single-stream text-to-image checkpoint from **Krea.ai, Inc.** — the full
classifier-free-guidance base model (and the LoRA-training base) behind the distilled Krea 2 Turbo.
This repository is a **Derivative** prepared for [`mlx-gen`](https://github.com/michaeltrefry/mlx-gen)
(and the SceneWorks worker that embeds it): the weights are group-wise-affine **quantized and repacked**
from the original bf16 diffusers checkpoint so the model loads and runs natively on Apple Silicon with no
Python/PyTorch sidecar. The `bf16/` tier is the dense original re-layout (max fidelity + the LoRA-training
base).

This is **not** the original checkpoint. For the reference model, training details, and the canonical
diffusers / SGLang inference paths, see the upstream card: **<https://huggingface.co/krea/Krea-2-Raw>**.

## Attribution

- **Original model:** Krea 2 Raw — © **Krea.ai, Inc.**, released 2026-06-22.
- **Base model:** [`krea/Krea-2-Raw`](https://huggingface.co/krea/Krea-2-Raw) (the undistilled base; Krea 2
  Turbo is distilled from it).
- **This Derivative:** quantized + MLX-repacked by the SceneWorks / `mlx-gen` project. No retraining or
  fine-tuning was performed — only numerical quantization and on-disk re-layout.

## License

Use of these weights is governed by the **Krea 2 Community License Agreement** and the Krea Acceptable Use
Policy, exactly as for the original model. A copy of the license is included in this repository as
[`LICENSE.pdf`](LICENSE.pdf) (also at
<https://huggingface.co/krea/Krea-2-Raw/blob/main/LICENSE.pdf>). In the event of any conflict, the Krea
Acceptable Use Policy and Krea 2 Community License control.

> **Deployer obligation (content filtering).** The Krea 2 Community License requires anyone who deploys the
> model to implement content-filtering measures or equivalent review processes appropriate to their use
> case, to prevent the generation or distribution of unlawful or policy-violating content. If you serve
> this model, you are responsible for those safeguards. Report harmful, illegal, or policy-violating
> outputs to **safety@krea.ai** (potential CSAM is escalated to NCMEC as required by law).

Krea does not claim copyright over generated outputs; users are solely responsible for their inputs and any
use of the outputs.

## What changed vs. the upstream checkpoint

The conversion is **lossy only through quantization** — the architecture, tokenizer, scheduler config, and
VAE are byte-for-byte the originals.

- **Transformer (DiT)** and **Qwen3-VL-4B text encoder**: for the Q8 / Q4 tiers the linear projection
  weights are quantized to **group-wise affine Q8 / Q4** (group size 64) and repacked into a single
  `.safetensors` per stack. Norms, embeddings, modulation tables, and the text-encoder vision tower stay
  dense. The `bf16/` tier keeps every weight dense.
- **VAE** (`AutoencoderKLQwenImage`): copied **unchanged** (f32).
- **`tokenizer/`, `scheduler/`, `model_index.json`**: copied unchanged.

## Repository layout

Each tier is a complete, self-contained snapshot you can load directly:

| Path   | Quantization       | On-disk size | Notes                                                        |
|--------|--------------------|--------------|--------------------------------------------------------------|
| `bf16/`| none (dense bf16)   | ~35.7 GB     | Max fidelity; the LoRA-training base.                        |
| `q8/`  | Q8 (group size 64)  | ~20.6 GB     | **Default.** Near-lossless; needs a 48 GB-class Mac.         |
| `q4/`  | Q4 (group size 64)  | ~12.5 GB     | Lighter footprint; mild quality trade-off.                  |

```
krea-2-raw-mlx/
├── LICENSE.pdf
├── README.md
├── bf16/ { transformer/  text_encoder/  vae/  tokenizer/  scheduler/  model_index.json }
├── q8/   { transformer/  text_encoder/  vae/  tokenizer/  scheduler/  model_index.json }
└── q4/   { transformer/  text_encoder/  vae/  tokenizer/  scheduler/  model_index.json }
```

## Usage

Built for Apple-Silicon inference through `mlx-gen`'s `krea_2_raw` engine. Point a loader at a tier
subdirectory (`bf16/`, `q8/`, or `q4/`); it auto-detects the packed weights. Unlike the CFG-free Turbo,
Krea 2 Raw is a **true classifier-free-guidance** model — run ~52 steps with a real guidance scale (~3.5)
and an optional negative prompt. The same `bf16/` tier is also the base for Krea 2 LoRA training.

## Model details

See the upstream card for the full model overview, capabilities, intended/out-of-scope uses, training-data
summary, safety measures, and risk/limitation disclosures: <https://huggingface.co/krea/Krea-2-Raw>.