Qwen-Image-2.1-Turbo: open 2K image gen in 8 steps, not 40
Same 7B weights as the base model, a hosted API near a cent per image, and a research-license catch—plus ElevenLabs voices 50% off on OpenRouter.
Copy markdown
8 steps, not 40, and the weights are public
Alibaba's new Turbo checkpoint runs the same 7B DiT (plus a Qwen3-VL 8B text encoder) as Qwen-Image-2.1 but cuts generation to 8 denoising steps from the base model's 40 — roughly 5x fewer — while keeping 2K output (up to 2048×2048) and RGBA transparency. The 8-step schedule is baked into the checkpoint and loads automatically via Diffusers' QwenImage21Pipeline; weights are on Hugging Face and ModelScope.
Build it this weekend: local 2K art, or ~1.4¢/image hosted
Community GGUF quants (Q4–Q8) already run in stable-diffusion.cpp and ComfyUI, and the base model fits in ~11GB with GGUF — so one consumer GPU can do 2K generation and multi-reference editing in 8 steps. Prefer hosted? Alibaba Cloud Model Studio serves qwen-image-2.1-turbo at about CNY 0.1 (~1.4¢) per image at 120 RPM, versus CNY 0.25 and 20 RPM for the Pro model.
The catch: downloadable isn't the same as commercial
Turbo ships under the Qwen Research License, not Apache — the same move the base 2.1 line made away from Apache 2.0. You can download and evaluate it freely, but commercial self-hosting needs separate permission from Qwen. If you need clean commercial rights, weigh the open-image alternatives, and read each model's license before you ship.
Cheaper voices this week: ElevenLabs 50% off on OpenRouter
OpenRouter is running ElevenLabs voice models — including Eleven v4, v4 Turbo and the v3 conversational line — at 50% off through Oct 19. If you route TTS or voice-agent traffic through OpenRouter, it's a short window to cut audio costs with no code change.