You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These weights are a converted version of Krea 2 Turbo and are governed by the Krea 2 Community License Agreement (v.1, 22 June 2026), a copy of which is in this repository. By requesting access you agree to be bound by it. In particular: commercial use is permitted only if your total company-wide annual revenue, across all affiliated entities and from all sources on a trailing twelve-month basis, is below US $1,000,000; you must implement reasonable content filtering in any deployment end users can reach; and if you redistribute these weights or anything containing them you must pass on the agreement, bind your recipients to it, name any derived model beginning with "Krea", and state that the model was modified.
Log in or Sign Up to review the conditions and access this model content.
Krea-2-Turbo-MLX
Krea 2 Turbo converted for Apple MLX, runnable by MLXBundle without Python at runtime.
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Modified: this is a modified version of Krea-2-Turbo. Not an official Krea product, and not endorsed by Krea.
Why this repository is gated
§3.1(a) of the agreement requires anyone distributing the model to provide a copy of it and require each recipient to be bound by it. A licence file in an open folder does the first and not the second. The gate does both.
What you are agreeing to
The full text is in LICENSE.pdf. The parts that bite:
Revenue threshold (§2.3). Commercial use is permitted only below US
$1,000,000 of total company-wide annual revenue - across all affiliated
entities under common ownership or control, on a trailing twelve-month basis,
from all sources, not only from the model. At or above that figure you need an
enterprise licence from Krea first: opensource@krea.ai. If you cross it
while using the model you must cease commercial use immediately.
Content filtering (§4.2). "You must implement reasonable and appropriate Content Filter measures to detect, prevent, and mitigate the generation or distribution of prohibited, harmful, or unlawful content through your deployment." Neither these weights nor MLXBundle include one. The agreement gives examples - open-source classifiers, commercial moderation APIs, human review - and disclaims endorsement of any of them.
If you redistribute (§3.1, §3.2). Pass on the agreement and bind your recipients to it; begin the name of any derived model with "Krea"; carry the attribution notice above; state that the model was modified; do not impose terms that conflict with the agreement or weaken Krea's warranty and liability protections; do not present your version as official or endorsed.
Other terms. Where law or platform policy requires it, disclose that output is AI-generated (§4.3). You indemnify Krea against claims arising from your use, your derivatives and your outputs, including claims about content and about missing filters (§8) - a one-way obligation. The licence terminates on any breach, on thirty days' notice at Krea's convenience, or if you bring a claim against Krea over the model (§9). Governing law is Delaware.
What conversion changed
- Tensor layouts transformed for MLX: the channel axis moved for convolutions, and the transposes each convolution kind needs.
- Weight-normalization parameter pairs fused into an equivalent inference representation.
- The transformer and text encoder quantized to 4 bits with a group size of 64.
Quantization is lossy - dequantized values approximate the originals.
manifest.jsonrecords which tensors were held back. - The VAE left in float32.
No training, fine-tuning, distillation or merging, and no learned parameters added. Verified against the reference implementation tensor by tensor: the transformer at 2.178e-07 on its output, the text encoder between 2.7e-07 and 3.9e-07, the VAE at 3.096e-05, with tiling reproducing the reference's own 2.810e-01 divergence from a single pass to four figures.
The bundle is self-contained at 9.81 GiB - transformer, encoder, VAE and tokenizer - where the bf16 original needs 39 GiB across two bundles and cannot be loaded at all on a 32 GB machine.
Using it
mlxbundle-cli krea ~/models/Krea-2-Turbo-MLX ~/models/Krea-2-Turbo-MLX \
"a red fox standing in deep snow at dusk" fox.png 512 512 8 0
Downloading a gated repository needs a Hugging Face token with read access; see the MLXBundle documentation.
Disclaimer
An independent conversion. No warranty of any kind, and no representation that any particular use of it complies with the agreement or with any law. The obligations above are yours. If your use is commercial, read the agreement rather than this summary.
Quantized