Consistency diffusion language models: Up to 14x faster inference without sacrificing quality
Together AI
Researchers introduced Consistency Diffusion Language Models (CDLM), which accelerates diffusion-based language model inference by combining consistency-based training with block-wise key-value caching. The method reduces refinement steps by 4.1x to 7.7x and achieves latency improvements up to 14.5x on coding benchmarks while maintaining quality through trajectory-consistent training objectives. This enables diffusion models to compete with autoregressive approaches on inference speed while preserving their bidirectional context capabilities.
Why it matters
Standard diffusion language models can't use KV caching and need too many refinement steps to be practical. CDLM fixes both with a post-training recipe that enables exact block-wise KV caching and trajectory-consistent step reduction — delivering up to 14.5x latency improvements
