Qwen-Image-2.1-Turbo GGUFs shrink the denoiser to 2.55 GB

AtomicChat’s GGUF release brings Qwen-Image-2.1-Turbo to stable-diffusion.cpp, reducing the 14.23 GB BF16 denoiser to between 2.55 GB and 7.59 GB. The repository includes perceptual-quality measurements for every quantization tier, giving developers concrete trade-offs among file size, memory use, and fidelity.

• Model: A 7B image-generation and editing checkpoint distilled for eight denoising steps.
• Quantizations: Q8_0 and model-aware variants from AD-Q6_K through AD-Q2_K.
• Smallest denoiser: 2.55 GB, excluding the text encoder, VAE, and runtime allocations.
• License: Research use under the Qwen Research License Agreement, with a separate agreement required for commercial use.

From 40 steps to eight

Alibaba’s Qwen team released Qwen-Image-2.1-Turbo as an accelerated checkpoint of the open-weight Qwen-Image-2.1 model. It generates and edits images in eight denoising steps, compared with the base model’s 40-step default, cutting denoiser passes by 80% while retaining the same 7B architecture.

Moving to eight steps does not guarantee a fivefold end-to-end speedup because text encoding, VAE decoding, memory transfers, and hardware utilization still contribute to runtime. Turbo also uses classifier-free guidance at CFG=1 by default and supports prefix KV caching, which reuses fixed text and reference-image context across denoising steps.

How Atomic Dynamic protects quality

Atomic Dynamic quantization, identified by the AD- prefix, assigns precision per tensor instead of applying one type uniformly. AtomicChat built an importance matrix from 32 Turbo renders whose prompts were excluded from the evaluation set, then allocated more precision to tensors that had a larger effect on output quality.