Consistency Diffusion Language Models

1 post · Feb 2026

What this trail covers

Together AI introduces Consistency Diffusion Language Models (CDLM), a post-training recipe that accelerates Diffusion Language Model (DLM) inference by combining consistency-based multi-token finalization with block-wise KV caching. This approach significantly reduces refinement steps and enables exact KV caching, leading to up to 14.5x latency speedups on math and coding tasks while maintaining competitive accuracy. CDLM allows for block-wise causal attention, improving hardware utilization and offering a balanced operating point for small-batch settings.

All Together AI trails

Feb 2026Consistency diffusion language models: Up to 14x faster inference without sacrificing qualityThis post introduces Consistency Diffusion Language Models (CDLM), a post-training recipe that accelerates DLM inference. CDLM combines consistency-based multi-token finalization with block-wise KV caching, enabling a block-wise causal student model that can reuse KV caches for finalized blocks. It employs three training objectives: distillation loss for newly unmasked positions, consistency loss for still-masked positions, and an auxiliary masked-denoising loss. CDLM achieves up to 14.5x latency speedups on math and coding tasks by reducing refinement steps and enabling exact KV caching, while maintaining competitive accuracy. The analysis also covers system-level performance characteristics, showing CDLM occupies an intermediate regime between AR and vanilla DLMs in terms of arithmetic intensity.
Advertisement