LLaDA2.0-Uni: One Diffusion Model for Understanding and Generating Text, Images, and Video

Inclusion AI’s 16B Mixture-of-Experts model uses masked diffusion instead of left-to-right generation, handling image understanding, text-to-image, and editing with only ~1B parameters active per token.
artificial-intelligence
Author

Kabui, Charles

Published

2026-04-28

Keywords

discrete-diffusion, multimodal-models, mixture-of-experts, image-generation, inclusion-ai

Download as PDF

Inclusion AI released LLaDA2.0-Uni, a multimodal model that uses discrete diffusion instead of autoregressive generation to both understand and create text, images, and video. The architecture combines three parts: a visual tokenizer (SigLIP-VQ) that converts images into discrete tokens identical to text tokens, a 16B Mixture-of-Experts backbone that processes all modalities through masked diffusion while activating only ~1B parameters per token, and a diffusion decoder for image reconstruction. Because text and images share one token space, the same model handles image captioning, text-to-image generation, and image editing through different masking patterns. A distilled decoder cuts image generation from 50 steps to 8, roughly 10× faster. The full pipeline needs about 47 GB of VRAM; understanding alone fits in 35 GB.

Most multimodal models bolt image generation onto a language backbone as a separate system. LLaDA2.0-Uni treats both as the same operation: denoise masked tokens. One model, one training objective, one inference stack handles everything from answering questions about a photo to generating new images from text. The Apache 2.0 license and full weight release make it accessible for any research team. Understanding-only mode fits on a single 40 GB GPU, keeping experimentation practical outside massive compute clusters.

Autoregressive generation has dominated LLMs since GPT-2. Continuous diffusion dominates image synthesis. LLaDA2.0-Uni, alongside Meituan’s LongCat-Next, represents a growing bet that discrete diffusion can unify both. If the approach scales, future foundation models may drop autoregression entirely, generating all tokens in parallel rather than one at a time.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {LLaDA2.0-Uni: {One} {Diffusion} {Model} for {Understanding}
    and {Generating} {Text,} {Images,} and {Video}},
  date = {2026-04-28},
  url = {https://toknow.ai/posts/llada-20-uni-discrete-diffusion-llm-multimodal-understanding-generation/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “LLaDA2.0-Uni: One Diffusion Model for Understanding and Generating Text, Images, and Video.” https://toknow.ai/posts/llada-20-uni-discrete-diffusion-llm-multimodal-understanding-generation/.