OpenAI Consistency Diffusion Language Models (CDLM) - Up to 14x Faster Inference

AI Impact Summary

This release introduces Consistency Diffusion Language Models (CDLM), a new training recipe that dramatically improves inference speed for diffusion language models by up to 14.5x without sacrificing quality. CDLM achieves this through block-wise KV caching and trajectory-consistent step reduction, enabling parallel generation and significantly reducing the number of refinement steps required. This is a capability update that will benefit applications requiring fast and efficient text generation, particularly those dealing with complex tasks like reasoning and math.