Diffusion LLM
Also known as: dLLM, diffusion language model, masked diffusion language model, discrete diffusion model
Standard large language models generate text autoregressively: they predict the next token, then the next, one at a time, in sequence. A diffusion LLM works differently. It starts with a sequence where all or many tokens are masked or noisy, then runs multiple denoising passes that progressively reveal the final output. The whole sequence is visible and can be refined in parallel across passes, which opens up different performance tradeoffs compared to the strict left-to-right dependency of autoregressive generation.
The approach adapts diffusion modeling, already widely used for image generation in tools like Stable Diffusion and Flux, to discrete text tokens instead of continuous pixel values. Research on diffusion LLMs accelerated significantly in 2025 and 2026, with papers from ByteDance, Google, and academic groups exploring architectures like masked diffusion and large-scale text diffusion. Google's DiffusionGemma work demonstrated real speed gains for text generation. The September 2026 frontier model dispatch called diffusion decoding one of five structural architecture trends redefining the stack.
For most builders, diffusion LLMs are not yet a practical deployment choice: the leading production models are still autoregressive transformers, and tooling support for diffusion-based text models is thinner. But the architecture is worth understanding because it changes the latency profile, cost structure, and parallelism of inference in ways that matter for future model selection. As the research matures and inference stacks add support, diffusion LLMs may become a real option for use cases that favor quality over token-by-token streaming.