LLaDA2.0-Uni: A Diffusion-based Large Language Model for Unified Multimodal Understanding and Generation

LLaDA2.0-Uni is a natively unified multimodal understanding and generation framework based on the discrete diffusion large language model architecture. It simultaneously achieves visual understanding and image generation in a single model, solving the problem of separated understanding and generation tasks in traditional multimodal systems and pioneering a new paradigm for next-generation foundation models.

Traditional multimodal systems adopt a composite architecture (language model + visual encoder + independent generation model), which has problems such as inconsistent representation spaces, split training objectives, and inability to natively support interleaved generation and reasoning. Most recent attempts are patchwork modifications to the dominant architecture, making it difficult to achieve true unification.

Natively building multimodal capabilities based on the discrete diffusion large language model (dLLM): 1. Fully semantic discrete tokenizer (text uses vocabulary embedding, images are discretized into semantic tokens via SigLIP-VQ to unify modal boundaries); 2. MoE-enhanced diffusion backbone (sparse activation adapts to multimodality, block-level masked diffusion unifies training objectives); 3. Diffusion decoder (few-step distillation optimization for fast pixel image reconstruction).

Data planning constructs large-scale datasets including image-text pairs, interleaved multimodal documents, and edit-generation datasets, focusing on semantic consistency. Training is divided into four stages: unimodal pre-training → multimodal alignment → capability integration → scenario fine-tuning.

The problem of slow inference in diffusion models is solved through prefix-aware optimization (direct encoding of the understanding task prefix, generating only partial diffusion) and few-step distillation decoder (compressing image generation steps to a few steps/single step).

Multimodal understanding benchmarks reach state-of-the-art levels; image generation follows complex prompts with precise and controllable editing; natively supports interleaved generation and reasoning, with a single model completing the understanding + generation process, supporting new interactions such as multi-turn dialogue and visual chain of thought.

Proves the feasibility of the unified architecture, challenging traditional cognition; open-source provides a research foundation; simplifies enterprise deployment and maintenance, reduces system complexity, and improves user experience.

Currently only supports image-text modalities, and inference efficiency still needs improvement; future directions: expand to video/audio temporal modalities, scale up the model to explore emergent capabilities, and enhance safety and controllability.