Traditional autoregressive models face a precision-vs-speed either-or dilemma — large models are slow, small models aren't accurate enough. Diffusion Language Models (dLLM) are breaking this deadlock.
The core idea of dLLM is to first generate a coarse text with placeholders, then use bidirectional attention to iteratively refine the entire text segment. Each iteration makes the output more accurate; the more iterations, the higher the quality. This means runtime can dynamically switch between latency and precision: voice assistant needs millisecond-level response? Use 2-3 steps. Complex code reasoning needs high quality? Use 20+ steps. Same model, no need to maintain multiple versions or complex routing logic.
The architecture has gone through three stages of rapid evolution. The first generation refined full context in parallel but couldn't use KV cache, with too high compute cost; the second generation introduced block-wise causal attention, doing local refinement in 8-64 token blocks, beginning to have practical value; the third generation continues optimizing token editing and streaming decoding capabilities.
For inference service providers and edge devices, dLLM means more flexible compute allocation strategies. The open-source community has released lightweight versions like LLaDA 2.0-mini, runnable on consumer-grade GPUs. Of course, dLLM is still in early stages; whether the extra latency from iterative refinement can be exchanged for sufficient precision gains needs broader validation. But when model architecture itself starts to break the binary opposition of big-and-slow vs small-and-fast, the efficiency curve of AI deployment will see significant change.