01
What inclusionAI released
On September 4, 2026, inclusionAI released runnable LLaDA-Image Base and LLaDA-Image-Turbo checkpoints together with Diffusers-based inference code. The official model zoo includes BF16 and FP8 checkpoints for both variants.
The release is best described as open weights plus inference code today. The project repository still marks training code as coming soon, so the live artifact release is not yet the same thing as a complete training-code release even though the paper publishes detailed training recipes.
02
What LLaDA-Image can do
The official repository says both Base and Turbo support text-to-image generation, VQ-conditioned generation, instruction-guided reference-image editing and Chinese-English text rendering. A single checkpoint family handles both generation and editing rather than requiring a separate editing model.
The paper and repository describe the family around a roughly 6B-parameter diffusion transformer with a frozen vision-language understanding module based on the LLaDA2.0-Mini backbone. For most users, the practical distinction is simpler: Base targets quality while Turbo targets low-step inference.
03
LLaDA-Image Base vs Turbo
The Base checkpoint is the quality-focused option and the official quick start recommends 50 sampling steps with guidance scale 5.0. LLaDA-Image-Turbo is a distilled variant using Twin-DMD and the official quick start recommends 4 steps with guidance scale 1.0.
04
Setup, software and hardware boundaries
The repository says its implementation has been used with Python 3.11, PyTorch 2.8, Transformers 4.57.6 and Diffusers 0.39.0. The documented path clones the repository, installs its requirements and runs the custom LLaDAImagePipeline on CUDA.
Text and VQ-conditioned generation dimensions must be divisible by 16, while editing dimensions must be divisible by 32. inclusionAI does not currently publish one universal minimum-VRAM figure that applies across Base, Turbo, BF16, FP8, resolution and generation mode, so fixed GPU-memory requirements should be treated as workload-specific measurements unless the project documents them directly.
05
Benchmarks and current limitations
inclusionAI reports overall Qwen-Image-Bench scores of 53.53 on the English track and 53.38 on the Chinese track and describes them as state of the art among open-source models on that benchmark. These are author-reported launch results, not independent validation or proof of universal superiority.
- Training code is still marked coming soon in the checked repository.
- No universal official VRAM minimum is published.
- Benchmark results are project-authored.
- A sufficiently clear commercial-use license was not exposed on the checked official release surfaces, so commercial deployment should wait for verified license terms.
06
Who should care about LLaDA-Image
Creators and developers now have a directly runnable generation-and-editing family with fast Turbo inference, Base quality mode and lower-precision checkpoint options. Agencies may find Turbo useful for concept iteration, but production use should be gated by license verification, measured hardware economics and workflow-specific quality testing.
For the practical install path, continue with the local-running guide covering Base, Turbo, BF16/FP8 choices, generation modes and editing without turning undocumented hardware estimates into official requirements.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.