Qwen-Image-2.1-Turbo
An accelerated open-source text-to-image model released by Alibaba's Qwen team, built on Qwen-Image-2.1. It applies distillation techniques to reduce the number of sampling steps, prioritizing faster inference speed. The model is suited for batch image generation and real-time interactive scenarios.
Timeline (last 90 days)
Alibaba's Qwen released Qwen-Image-2.1-Turbo, an 8-step 7B open-source image model, pointing to a latency-optimized release.
Unsloth v0.1.906 integrates the official 8-step version of Qwen-Image-2.1, Qwen-Image-2.1-Turbo, with a model size of 4.2 GB in Q4 GGUF, and 8-step sampling increases image generation speed by about 2.5× compared with the regular version.
Step compression is typically accompanied by a certain degree of quality trade-off, and whether 8-step generation can approach the 40-step version in complex compositions or fine textures still requires real-world testing to verify
Qwen-Image-2.1-Turbo requires only 8 steps to complete image generation and editing, while the base model defaults to a 40-step denoising process
Qwen-Image-2.1-Turbo maintains the same 7B parameter architecture as the base model, with the number of denoising steps reduced by 5×
Qwen-Image-2.1-Turbo is essentially a distillation-based acceleration of the original model
Qwen-Image-2.1-Turbo offers the option to access an officially hosted API
It is recommended that teams conduct a comparative evaluation based on their own scenarios before deciding whether to fully switch to the Turbo version or use both models on demand
Running Turbo on an M1 Pro 16GB at 1024² takes approximately 120 seconds.
Setting num_inference_steps alone does not override the schedule already saved by Turbo; the sigmas parameter must be explicitly passed in to change the number of steps.
16 more timeline events
All Facts (20)
Qwen-Image-2.1-Turbo maintains the same 7B parameter architecture as the base model, with the number of denoising steps reduced by 5×
50%UnverifiedIt is recommended that teams conduct a comparative evaluation based on their own scenarios before deciding whether to fully switch to the Turbo version or use both models on demand
50%UnverifiedStep compression is typically accompanied by a certain degree of quality trade-off, and whether 8-step generation can approach the 40-step version in complex compositions or fine textures still requires real-world testing to verify
50%UnverifiedQwen-Image-2.1-Turbo is essentially a distillation-based acceleration of the original model
50%UnverifiedQwen-Image-2.1-Turbo offers the option to access an officially hosted API
50%UnverifiedQwen-Image-2.1-Turbo requires only 8 steps to complete image generation and editing, while the base model defaults to a 40-step denoising process
50%UnverifiedThe Hugging Face repository for Qwen-Image-2.1-Turbo is 32.4 GB in size.
50%UnverifiedIn ComfyUI, using a regular KSampler at 8 steps cannot reproduce the Turbo schedule; SamplerCustomAdvanced with a ManualSigmas node must be used instead, with the sampler set to euler and CFG set to 1.
50%UnverifiedQwen-Image-2.1-Turbo completes image generation and editing using 8 denoising steps, whereas the base model requires 40 steps by default.
50%UnverifiedTurbo and the base version use the same 7B visual generation architecture and are loaded with the same QwenImage21Pipeline.
50%UnverifiedTurbo defaults to CFG=1, requiring only one forward inference pass per step.
50%UnverifiedQwen has not published any speed data or quality benchmarks for Turbo.
50%UnverifiedRunning Turbo on an M1 Pro 16GB at 1024² takes approximately 120 seconds.
50%UnverifiedThe actual parameter count of the full pipeline is 16.22B, comprising 7.115B for the Transformer, 8.767B for the text encoder (Qwen3-VL 8B), and 0.338B for the VAE.
50%UnverifiedSetting num_inference_steps alone does not override the schedule already saved by Turbo; the sigmas parameter must be explicitly passed in to change the number of steps.
50%UnverifiedThe license for Qwen-Image-2.1-Turbo excludes any commercial paid use.
50%UnverifiedThe official release of Turbo weights directly spares users from having to assemble accelerated LoRA components themselves.
50%UnverifiedThe specific differences between Qwen-Image-2.1-Turbo and the original 2.1 have not yet received an authoritative and detailed official explanation as of now.
50%UnverifiedFor developers who need to generate images in large batches or build interactive applications, the time savings brought by the Turbo version are more valuable.
50%UnverifiedAlibaba's Qwen released Qwen-Image-2.1-Turbo, an 8-step 7B open-source image model, pointing to a latency-optimized release.
50%