LLaDA-Image-Turbo is Ant Group's open-source image generator and editor for developers who need high-quality visuals without the latency of multi-step diffusion. Released by the company's BaiLing lab on September 4, 2026, it packages a roughly 6-billion-parameter unified architecture into two checkpoints: a Base model for maximum fidelity and a Turbo variant distilled to just 2 to 4 sampling steps. Both ship under open weights on Hugging Face and ModelScope, with BF16 and FP8 precision builds for different hardware budgets.
Core Features
- A single checkpoint handles text-to-image, reference-image editing, and instruction-guided edits with no separate editing backbone.
- Turbo uses Twin-DMD distillation to cut sampling to 2 to 4 steps, suiting real-time and batch production.
- Strong bilingual text rendering: it scores 53.53 on the English and 53.38 on the Chinese Qwen-Image-Bench, the current state of the art.
- Native Chinese and English prompt understanding, useful for mixed-language posters and product shots.
- Weights released openly so teams can self-host and avoid per-image API fees.
Use Cases
- Chinese-market creative teams rendering accurate in-image text on posters and ads.
- Indie developers building low-latency image features on consumer GPUs.
- Studios wanting a self-hosted alternative to closed image APIs.
Pricing
Free and open-weight under a permissive release, with no per-image fee. Hosting costs are your own GPU time. Enterprise support is arranged through Ant's BaiLing lab.
Our Take
Best for teams that need fast, bilingual image generation they can own and run themselves; the trade-off is that, like most open models, you supply the serving infrastructure and the Base checkpoint is slower than Turbo. Browse more open image models in our Image Generation category.
