Access Krea 2 Turbo iQ4.5
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By clicking "Agree", you agree to the Krea 2 Community License Agreement and acknowledge the Acceptable Use Policy.
Log in or Sign Up to review the conditions and access this model content.
Krea 2 Turbo iQ4.5
MLX only. An imatrix-calibrated conversion of krea/Krea-2-Turbo for mlx-serve on Apple Silicon. The weights are MLX's packed affine format with a trimmed text encoder, so diffusers, ComfyUI and GGUF runtimes cannot load them. This is a community quantization, not an official Krea product, and it is not endorsed by Krea.
Requires the next mlx-serve release, the first with calibrated image packs. Earlier builds refuse this pack at load: its text encoder keeps only the 35 layers Krea 2 reads, and earlier MLX Core apps look for a differently named transformer file.
| this pack | BF16 release | |
|---|---|---|
| total on disk | 9.94 GB | 35.7 GB |
| transformer | 7.21 GB at 4.5 bits per weight | 26.3 GB |
| text encoder | 2.21 GB at 4.5 bits per weight, 35 of 36 layers, no vision tower | 8.9 GB |
| VAE | 0.51 GB, unchanged | 0.51 GB |
Bits per weight count the stored scales and biases, so 4.5 is the size of plain 4-bit
group-64 quantization. For comparison, ddalcu/Krea-2-Turbo-MLX-Serve-mixed-4-8 is 15.83 GB.
How it was made
- Calibration from the engine itself. Per-input-channel activation statistics were collected inside mlx-serve while it generated 24 varied prompts (portraits, landscapes, illustration styles, product shots, lettering) across four aspect ratios at 8 steps.
- Weighted scale search. Each quantization group's scale and bias come from an activation-weighted least-squares search, with min/max only as the starting point.
- Measured bit allocation. Every linear's error was measured at 3, 4, 5, 6 and 8 bits, and the transformer's widths (3-bit x49, 4-bit x127, 5-bit x74, 6-bit x6, bf16 x8) were bought by error saved per byte inside the budget.
- Conditioning tensors kept at bf16. The input and output projections (
first,last.linear), the timestep MLP and its modulation projection (tmlp.0,tmlp.2,tproj.1), the text MLP (txtmlp.1,txtmlp.3) and the text-layer projector (txtfusion.projector). - Text-encoder MLP outputs kept wider.
down_projin layers 0-3, 6, 9, 12, 15, 18, 21, 24, 27 and 30-34 is held at 6-bit or wider, the layers llama.cpp's Q4_K_M widens. - Only the text encoder that runs. Krea 2 reads the Qwen3-VL language stack up to the input of layer 35, so layer 35, the final norm and the vision tower are not shipped.
- VAE untouched. No AWQ-style scale folding, so runtime LoRAs still compose with the stored weights.
On 8 held-out prompts at 768x768 and 8 steps it rendered close to a higher-precision (5.5/5.0 bits per weight) conversion of the same model: SSIM 0.68 mean.
Use
mlx-serve --serve --model /path/to/Krea-2-Turbo-iQ4.5 --port 11434
curl -s http://127.0.0.1:11434/v1/images/generations \
-H 'Content-Type: application/json' \
-d '{"prompt": "a lighthouse on a rocky coast at dusk", "size": "1024x1024", "steps": 8}'
License
Krea 2 is licensed under the Krea 2 Community License Agreement, included here as
LICENSE.pdf; the required attribution is in NOTICE. For more information, visit
https://krea.ai/krea-2-licensing. By downloading or using this model you are bound by that
agreement and the Acceptable Use Policy. In
particular:
- Content filtering (section 4.2). If you deploy this model you must add reasonable content filtering. mlx-serve does not include one.
- Commercial use (section 2.3). Allowed only if your organization's annual revenue is under US$1,000,000; above that you need an enterprise license from Krea (opensource@krea.ai).
- Downloads last month
- 5
Quantized