AIBRIEF
The 6 billion parameter image model from inclusionAI, an independent AI research group, leads a third party public image generation benchmark on both English and Chinese tracks, then releases its weights, training code, and recipe.
A new 6-billion-parameter image-generation model from inclusionAI, LLaDA-Image, has set the top open-source score on Qwen-Image-Bench, a public image-generation test, on both its English (53.53) and Chinese (53.38) tracks. Unlike most top-scoring models, the team has published not just the weights but the full training code and recipe.
The model pairs a from-scratch image-generation architecture with a frozen language-understanding module built on the team's earlier LLaDA2.0-Mini model. Training runs in two stages, image-only pre-training and mid-training, then paired image-text data, on a roughly 220-million-sample pipeline.
A distilled variant, LLaDA-Image-Turbo, runs in 2 to 4 sampling steps, the number of passes the model makes per image. That shortens time-to-image enough for live use.
The scores are paper-reported on a single benchmark, not an independent reproduction. The SOTA claim is inclusionAI's until a community leaderboard confirms it. That same openness lowers the bar for misuse, a real cost of any full image-generation release.