LLaDA-Image ships a 4-step open image model

Ant Group’s inclusionAI released LLaDA-Image and its 4-step Turbo checkpoint, a 6B model for image generation and editing; the base uses 50 sampling steps, while Turbo is distilled to 4, with reported Qwen-Image-Bench scores of 53.53 in English and 53.38 in Chinese. The repository ships weights and inference code but marks training code as coming soon, so builders should test the four-step path against their own latency and quality budget before treating it as a production default. That makes it a useful new datapoint for open-weight models moving from research novelty into deployment economics.