Running Krea 2 Turbo natively on Mac (Swift/Metal, no Python
Hi! Krea 2 Turbo is now in Siliconed, a free, open-source macOS app that runs image models natively on Apple silicon, in Swift and Metal: no Python, no ComfyUI, nothing to wire. Pick Krea 2 in the app, choose an install, it downloads the weights, then you type a prompt. Text-to-image in the usual formats, and LoRAs (ai-toolkit layout).
- On a 16 GB laptop: on an M1 Pro, 8 steps at CFG 1, the full bf16 weights: ~146 s at 1024², ~217 s at 1024×1536 with Faster Rendering (on by default: the first 4 of the 8 evaluations on a half-size grid; ~221 s and ~355 s with all of them at full size).
- Little memory: the 34 GB of weights are streamed from disk, and the memory need is computed from the model's shapes before the first step. Peak at 1024²: 3.1 GB in Standard (bf16), 2.2 GB in Compact (int8 convrot), 1.4 GB in Light (GGUF Q4_K_M).
- Checked against the reference: each stage (Qwen3-VL-4B and its layer fusion, the DiT, the VAE) is compared channel by channel with diffusers in fp32; at 1024² the DiT output is 1.4e-4 from it. The Compact and the Light are checked to the bit against their reference dequantization, LoRAs against diffusers' and PEFT's merge.
- No sign-in needed: the app fetches open copies proven identical to Krea's files (Comfy-Org's bf16 and int8 weights, the Qwen-Image VAE), so it works without a Hugging Face account.
The app is Apache-2.0; the model keeps its Krea 2 Community license: the app shows it in full, with the attribution it asks for, before the first byte is downloaded. Notarized DMG:
https://gwenn-ha-dev.github.io/Siliconed
GH https://github.com/gwenn-ha-dev/Siliconed
[image: krea2.jpg — Krea 2 Turbo, Standard, 1248×832, 8 steps, seed 1001, on an M1 Pro]
I'm the author; happy to answer questions. Thanks to Krea for the model, to Comfy-Org for the open copies, and to vantagewithai for the GGUF!
