Krea 2 App Quantized

Public application artifacts for the native Swift/MLX Krea 2 app. This repository contains verified, unmodified official Krea shared dependencies and custom Q8 and Q6 quantized derivatives of the Krea 2 Turbo transformer. The Identity Edit LoRA remains in its official upstream repository at the immutable revision recorded in MODEL_MANIFEST.json.

These files are unofficial derivatives. They are not official Krea products and are not affiliated with or endorsed by Krea.ai, Inc.

Provenance

  • Base model: krea/Krea-2-Turbo
  • Base revision: 98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7
  • Source DiT file: turbo.safetensors
  • Source DiT SHA-256: 78bbf8f4165eda19cea3cb06c78089221932a39e2eed8af9da741f942c47ffb3
  • Q8 conversion: affine 8-bit, group size 64, krea-swift-q8-converter/1.0.0
  • Q6 conversion: affine 6-bit, group size 64, krea-swift-q6-converter/1.0.0
  • Conversion implementation checkpoint: krea-swift@10c9d1c91cd490d8d8b5bd724dd2c77d73e1ac16

Each tier contains 28 independently loadable DiT block shards, three dense shell shards, and the app's sealed quantized-artifact manifest. The manifest records tensor layout, exclusions, conversion parameters, source identity, per-shard sizes, and per-shard SHA-256 values.

Layout

shared/
  encoders/qwen/model.safetensors
  tokenizer/{chat_template.jinja,tokenizer.json,tokenizer_config.json}
  vae/{config.json,diffusion_pytorch_model.safetensors}
bf16/                                  # reserved; official file not yet hosted
q8/
q6/
LICENSE.pdf
NOTICE
README.md
MODEL_MANIFEST.json

License and use requirements

The Krea 2 model weights and these quantized derivatives are licensed under the Krea 2 Community License Agreement v1 dated June 22, 2026, not Apache-2.0. The complete agreement is included as LICENSE.pdf.

By downloading, accessing, copying, using, or distributing these artifacts, you acknowledge that you have read, understood, and agree to be bound by that agreement. You must also comply with Krea's Acceptable Use Policy: https://www.krea.ai/krea-2-use-policy.

Among other terms, the agreement requires downstream distributors to provide the agreement and require recipients to be bound by it; deployments must use reasonable and appropriate content filtering; commercial use under the community license is limited by its company-wide annual-revenue threshold; and applicable AI provenance/disclosure requirements must be followed. Review the complete agreement and policy rather than relying on this summary.

No additional license is granted here. These artifacts are provided subject to the warranties, liability limitations, restrictions, and other protections in the Krea 2 Community License Agreement.

Official and derivative files

Files under shared/ are byte-identical, unmodified official files distributed by krea/Krea-2-Turbo at revision 98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7. Files under q8/ and q6/ are modified quantized derivatives as described above.

The official BF16 turbo.safetensors is catalogued at the reserved path bf16/turbo.safetensors, but it is not currently hosted because its previously verified local source is no longer installed. It has not been downloaded, reconstructed, or substituted. Identity Edit remains external at conradlocke/krea2-identity-edit@89e9e7a09ee2e5c9331e952063d79b1b8a703280.

Exact paths, byte sizes, and SHA-256 values are recorded in MODEL_MANIFEST.json.

Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for stevenjmotion/Krea-2-App-Quantized

Base model

krea/Krea-2-Raw
Finetuned
(39)
this model