Qwen-Image-2.1 Turbo is the fast version of Alibaba's Qwen-Image-2.1, and the whole story fits in two numbers: 40 and 8. The base model's official examples use 40 denoising steps; Turbo ships with an 8-step schedule baked in. If you searched for qwen image 2.1 turbo, that is what you get: the same 7B model, sampling in a fifth of the steps.
Turbo and the hosted Pro/Turbo APIs went live on October 9, 2026. On Alibaba Cloud Model Studio it is qwen-image-2.1-turbo; the weights are Qwen/Qwen-Image-2.1-Turbo under the Qwen Research License Agreement.
Architecture and Sampling Steps
Nothing was shrunk. The visual generator is still the 7B, 32-layer Single-Stream DiT used by Qwen-Image-2.1, and text-to-image and editing share one Diffusers pipeline, QwenImage21Pipeline. The model card and the repository cover the rest:
- The 8-step schedule loads automatically, and setting
num_inference_stepsalone won't override it. - CFG defaults to 1; prefix KV caching reuses text and reference-image context across steps.
- Output can be transparent RGBA, editing takes up to 10 reference images, and aspect-ratio presets match the base model (2048×2048 at 1:1).
Alibaba doesn't publish a wall-clock speedup. Five times fewer steps is a ratio between official examples; your latency depends on GPU, resolution, caching and hosted versus local.
Qwen-Image-2.1 Base Model vs Turbo
The base checkpoint, Qwen/Qwen-Image-2.1, came first. Qwen released it on September 20, 2026, with day-0 support in Diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V and Intel's OpenVINO. Turbo followed on October 9. Both are on Hugging Face and ModelScope under the same Qwen Research License.
| Qwen-Image-2.1 (base) | Qwen-Image-2.1 Turbo | |
|---|---|---|
| Weights | Qwen/Qwen-Image-2.1 |
Qwen/Qwen-Image-2.1-Turbo |
| Released | Sep 20, 2026 | Oct 9, 2026 |
| Visual generator | 7B, 32 Single-Stream DiT layers | Same |
| Steps in official examples | 40 (num_inference_steps=40) |
8, saved in the checkpoint |
| Tasks | Text-to-image, editing with up to 10 references, transparent RGBA | Same |
| Hosted API on Model Studio | None under the base name; the closest hosted tier is Qwen-Image-2.1 Pro, qwen-image-2.1-pro |
qwen-image-2.1-turbo |
| Hugging Face downloads, last 30 days (Oct 22) | 176,590 | 30,766 |
The base model is the one to fine-tune or to run when you want full control over the step count and guidance. Its card lists seven aspect-ratio presets, from 2048×2048 at 1:1 to 2752×1536 at 16:9, and recommends enable_model_cpu_offload() for GPUs with less memory. Qwen also published prompt-enhancement checkpoints, Qwen-Image-2.1-PE-T2I and Qwen-Image-2.1-PE-I2I.
On Model Studio, the Pro API costs $0.04 per image in Singapore and $0.035333 in Beijing, Virginia, Frankfurt, Tokyo and Hong Kong, with a 20 RPM limit. Turbo costs less and allows six times the request rate, as the next section shows.
Per-Image Pricing on Model Studio
Alibaba's Turbo pricing page, updated October 14, 2026, excluding promotions:
| Region | Per image | Per 1,000 |
|---|---|---|
| Singapore (International) | $0.016 | $16 |
| Beijing, Virginia, Frankfurt, Tokyo, Hong Kong | $0.014133 | ~$14.13 |
The rate limit is 120 RPM in those regions, and any promotional price appears only in the console. The model info page calls Turbo the most balanced, cost-effective open-source model in the Qwen-Image line; it supports fine-tuning but not batch inference, function calling or web search. Self-hosting has no list price, and the research license deserves a careful read before any commercial launch.
Usage Data
AI Rank's OpenRouter image ranking for October 7–13, 2026 (by requests) shows qwen/qwen-image-3 and qwen/qwen-image-3-pro around No. 29 and 30. Qwen-Image-2.1 Turbo isn't on it that week, which says something about OpenRouter traffic and nothing about image quality. Alibaba hasn't published FID or CLIP comparisons between Turbo and the base model either.
→ Live image usage: See this week's order in the OpenRouter image model ranking on AI Rank (requests, Oct 7–13, 2026 window).
As an image model it has no Arena text page; the open-source rankings on AI Rank track GitHub growth for AI repos.
Compare before you choose: image model usage moves weekly — check the current OpenRouter image rankings on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 15, 2026; base-model and Pro API details added on October 22, 2026 from the Qwen-Image-2.1 model card, GitHub README, Hugging Face API and Model Studio's Pro page. We did not run our own benchmarks for this article.