Found first: a primary source the press has not covered yet.
A 30-author team led by Chuyan Chen has released LLaDA-Image, a 6B Diffusion Transformer image generator that scores 53.53 on the English track and 53.38 on the Chinese track of Qwen-Image-Bench. The authors describe both as state-of-the-art among open-source models. Model weights, training code, and detailed recipes are all public under CC BY-SA 4.0.
What the source says
LLaDA-Image pairs a 6B DiT trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Training ran across 220M samples, with an image-only pre-training phase before paired image-text data was introduced. The architecture uses parameter-free RMSNorm throughout the DiT and the Muon optimizer. A distilled variant, LLaDA-Image-Turbo, cuts inference to 2 to 4 sampling steps.
Why it matters
Open image generation releases often include weights but omit training recipes, making benchmarks hard to reproduce independently. LLaDA-Image publishes weights, training code, and data staging details together. The mid-training disclosure, covering the image-only pre-training phase before paired data was introduced, is specifically what enables independent reproduction of the 53.53 Qwen-Image-Bench score. At 6B parameters, the model is large enough to be competitive but small enough for teams without frontier-scale compute to work with.