Downloads · 30 days
1.3K
36% of all-time downloads
mh73772/krea-2-turbo-mflux-q8-4step
krea-2-turbo-mflux-q8-4step is a text-to-image model from mh73772. Use it when you need an image from a text prompt. It is set up for mflux. The card lists the license as other.
An 8-bit MLX checkpoint of krea/Krea-2-Turbo, prepared to run on a 16 GB Apple Silicon Mac with the --block-streaming mode of the mflux-for-16gb fork.
Downloads · 30 days
1.3K
36% of all-time downloads
All-time downloads
3.6K
Public
Repo size
18.4 GB
Likes
0
Public
Click a slice to open those files.
.safetensors18.4 GB · 100%
From the Hugging Face model README
An 8-bit MLX checkpoint of krea/Krea-2-Turbo, prepared to run
on a 16 GB Apple Silicon Mac with the --block-streaming mode of the
mflux-for-16gb fork.
Not an official Krea product. See NOTICE and LICENSE.pdf.
mflux-community/krea-2-turbo-mflux-q8| Directory | Size | Change |
|---|---|---|
transformer/ | 13.62 GB | lvladikov/Krea2-Turbo-Distill-4step-LoRA (krea2_turbo_4step_rank_64_lora_comfyui.safetensors, scale 1.0) merged into the q8 weights, then re-split into blocks_00–blocks_27.safetensors + globals.safetensors so each block can be read from disk on its own |
text_encoder/ | 4.27 GB | Qwen3-VL-4B quantized from bf16 (8.05 GB) to 8-bit, MLX affine, group size 64 |
vae/, tokenizer/ | 0.52 GB | unchanged |
The 4-step LoRA is already merged: do not apply it again.
Requires the fork above (--block-streaming is not in upstream mflux).
hf download mh73772/krea-2-turbo-mflux-q8-4step --local-dir krea-2-turbo-mflux-q8-4step
mflux-generate-krea2 \
--model ./krea-2-turbo-mflux-q8-4step --base-model krea-2 \
--block-streaming \
--steps 4 --scheduler euler --guidance 1.0 \
--width 1024 --height 1024 --seed 42 \
--prompt "..." --output out.png
| Resolution | Time per step | Peak phys_footprint | Swapouts |
|---|---|---|---|
| 1024² | 7.90 s | 6.18 GB | 0 |
| 1280² | 14.36 s | 6.53 GB | 0 |
4 steps, euler, guidance 1.0. Measurements and method: docs/16gb/ in the fork.