# Qwen Image 2.1 Turbo Qwen Image 2.1 Turbo is the 8-step checkpoint of Qwen-Image-2.1, the image generation and editing model from [[Qwen|Alibaba's Qwen team]]. It landed on [[HuggingFace]] on 9 October 2026, about four weeks after the full Qwen-Image-2.1 (14 September 2026). Same 7B visual generation architecture, far fewer sampling steps. Why should you care? After [[Qwen Image 3.0]] shipped without any weights, this is the Qwen image model you can actually download and run on your own machine. And about ten hours after the official upload, a community GGUF quant appeared that fits on a 6-8 GB GPU. ## What Qwen-Image-2.1 brings Turbo inherits all of this from the base model: - **7B parameters** in the visual generation part (32 single-stream DiT layers), the same size class as [[Qwen Image 2.0]] - **Native transparency**: it generates RGBA images, edits transparent layers, and cuts subjects out of photos - **Creation and editing in one model**, with up to 10 reference images and local edits marked by circles, painted annotations or separate masks - **Better typography, portrait lighting and fine detail**, according to Qwen (the card has no independent benchmark) - **Around 2K output**: 2048×2048 for 1:1, up to 2752×1536 for 16:9 ## What "Turbo" changes Qwen calls it an "accelerated checkpoint"; the quant's card calls it distilled. Same model size, but it's trained to get there in 8 denoising steps instead of the 40 used in the base model's example. Think of it as a cousin of [[Knowledge Distillation]], where the thing you shrink is the number of steps rather than the model. - The 8-step schedule ships inside the checkpoint and loads automatically. Setting `num_inference_steps` alone does NOT override it - CFG (classifier-free guidance, see [[Diffusion Models]]) is fixed at 1, so negative prompts are ignored - Prefix KV caching reuses the text and reference-image context across denoising steps - It runs with `QwenImage21Pipeline` in Diffusers, which currently means installing Diffusers from source ## The GGUF quant by Abiray `Abiray/Qwen-Image-2.1-Turbo-GGUF` is a community conversion, NOT an official Qwen release. Abiray is a busy Hugging Face quantizer who also publishes GGUF builds of LTX, Wan 2.2, MiniMax, Gemma and the Qwen LLMs. These files were converted with the ComfyUI-GGUF tools and quantized with a patched `llama-quantize` from [[llama.cpp]] that understands the 297-tensor `qwen_image` architecture. [[GPT-Generated Unified Format (GGUF)|GGUF]] gives you the usual [[AI Quantization]] trade: smaller weights and less VRAM, paid for with some quality at the low end. | Quant | File size | VRAM (card's advice) | Notes from the card | |---|---|---|---| | Q8_0 | 7.59 GB | 12-16 GB+ | "Indistinguishable from native BF16" | | Q6_K | 5.88 GB | 10-12 GB | Keeps complex typography and prompt nuance | | Q5_K_M, Q5_K_S, Q5_0 | 4.94-5.01 GB | 8-12 GB | | | Q4_K_M | 4.19 GB | 6-8 GB | The recommended pick | | Q4_K_S, Q4_0 | 4.05-4.06 GB | 6-8 GB | | | Q3_K_M, Q3_K_S | 3.11-3.19 GB | 4-6 GB | Laptops and tight memory | | Q2_K | 2.45 GB | 4 GB | Noticeable texture loss | Read that table with two things in mind. First, the VRAM column is the quantizer's advice, not a measurement. Second, the `.gguf` is only the diffusion model: you still need the text encoder (a Qwen3-VL 8B, in int8 or w4a8) and the VAE, both from `Comfy-Org/Qwen-Image-2.1`, and they add to your memory budget. ## How to run it The card documents one path, [[ComfyUI]]: 1. Install the ComfyUI-GGUF custom node (by city96) 2. Put the `.gguf` in `models/diffusion_models/` (or `models/unet/`), the text encoder in `models/text_encoders/` and the VAE in `models/vae/` 3. Drag in the text-to-image or image editing workflow JSON linked from the card, then point the loader at your quant 4. Use the `euler` sampler, the `simple` or `normal` scheduler, CFG 1.0 and an empty negative prompt One inconsistency is worth knowing about. The quant card says 4 to 8 steps, with 6 as "ideal". Qwen's own card says the checkpoint was tested with its saved 8-step schedule and that other schedules haven't been evaluated. I'd start at 8 and only drop lower when speed matters more than detail. The card says nothing about stable-diffusion.cpp or other runtimes. As of October 2026, ComfyUI is the documented route. ## License The license is the Qwen Research License (`qwen-research`, dated 20 September 2026). It allows research and evaluation only; commercial use requires a separate license from Alibaba. The GGUF quant inherits the same terms. That's a step back. Qwen-Image 1.0 and [[Qwen Image 2.0]] shipped under Apache 2.0. The 2.1 weights are downloadable, but you can't build a product on them without asking first. If you need commercial use and few steps, [[FLUX.2]] Klein 4B is Apache 2.0 and runs in 4 steps. ## My take For local experiments, this is a really attractive package: editing, RGBA output and multi-reference composition in a 7B model, 8 steps per image, and a Q4 file of about 4 GB. I'd use it to explore and prototype. For client work or anything you sell, the license rules it out, so look at [[FLUX.2]] Klein or [[Stable Diffusion]] instead. ## References - GGUF quant (Abiray): https://huggingface.co/Abiray/Qwen-Image-2.1-Turbo-GGUF - Official Turbo checkpoint: https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo - Qwen-Image-2.1 base model: https://huggingface.co/Qwen/Qwen-Image-2.1 - Qwen-Image-2.1 blog: https://qwen.ai/blog?id=qwen-image-2.1 - GitHub: https://github.com/QwenLM/Qwen-Image-2.1 - License: https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/main/LICENSE - ComfyUI-GGUF: https://github.com/city96/ComfyUI-GGUF - Text encoder and VAE for ComfyUI: https://huggingface.co/Comfy-Org/Qwen-Image-2.1 ## Related - [[Qwen]] - [[Qwen Image 2.0]] - [[Qwen Image 3.0]] - [[AI image generation models]] - [[AI Image Generation (MoC)]] - [[Diffusion Models]] - [[ComfyUI]] - [[GPT-Generated Unified Format (GGUF)]] - [[AI Quantization]] - [[Running AI Models Locally]] - [[AI Open Weight Models]] - [[HuggingFace]] - [[llama.cpp]] - [[FLUX.2]] - [[Stable Diffusion]] - [[Knowledge Distillation]]