AtomicChat publishes quantized GGUF weights for Qwen-Image-2.1-Turbo
AtomicChat has published GGUF quantizations of Qwen-Image-2.1-Turbo for use with stable-diffusion.cpp. The model card describes Turbo as a 7B image generator distilled for eight denoising steps at CFG 1, with the same text encoder and VAE as Qwen-Image-2.1.
The card reports paired comparisons on 48 prompts and two seeds against BF16 outputs. Its AD-Q4_K file is 4.20 GB; the authors report it is 21% closer to BF16 than plain Q4_K, with 150 MB of additional size. They also publish a fixed sampling schedule, settings, and measurement details. These are measurements reported by the repository authors, not an independent evaluation. The card says Turbo requires its own schedule; using the base model's default schedule changes the output.