Ant Group’s open-source lab inclusionAI released the LLaDA-Image checkpoints on September 4, 2026, with inference code and weights for a 6 billion parameter family that does text-to-image generation and instruction-guided editing from one checkpoint. The model card lists a 50-step Base model and LLaDA-Image-Turbo, a Twin-DMD distilled variant meant to generate or edit in 2 to 4 steps.

The same day the lab pointed readers at the paper, LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes, posted to arXiv on September 3. The authors describe a unified diffusion stack in which both the backbone and the DiT are diffusion models, trained first with image-only pre-training before paired language supervision and joint generation-editing work. Both variants handle text-to-image, VQ-conditioned generation, reference-image editing, and Chinese and English text rendering.

inclusionAI, backed by Ant Group, said LLaDA-Image posted overall Qwen-Image-Bench scores of 53.53 in English and 53.38 in Chinese. Weights are on Hugging Face and ModelScope in BF16 and FP8. Training code is listed as coming soon. Recommended settings on the card are 50 steps and guidance 5.0 for Base, and 4 steps and guidance 1.0 for Turbo, at 1024 by 1024.

The GitHub repo at inclusionAI/LLaDA-Image ships Diffusers-style inference for Python 3.11, PyTorch 2.8, Transformers 4.57.6, and Diffusers 0.39.0. The card is explicit that this is a unified generator and editor, not a generator plus a separate editing backbone.

Decoded Take

Western labs still treat training recipes as the moat. inclusionAI is publishing the weights, the sampling code, and a paper that starts from image-only pre-training, which is the part most “open” image models skip. A 4-step 6 billion parameter editor that holds up on bilingual text is enough to put pressure on closed APIs that charge by the megapixel. The unfinished half is the training code. Until that lands, this is an open checkpoint with a closed kitchen, and every reproduction argument will stall there. Watch whether the training release shows up before the next inclusionAI drop, whether independent benches confirm those 53-point Qwen-Image-Bench numbers, and whether Ant keeps this under a permissive license once the model is good enough for commercial posters.