DFlash claims up to 15x faster LLM inference on Nvidia Blackwell

MarkTechPost reported that DFlash is a speculative-decoding approach that drafts whole token blocks in parallel for LLM inference. The method uses block diffusion drafting and KV injection and is reported to deliver up to 15x higher throughput on Nvidia Blackwell. The work targets the inference bottleneck that matters most for interactive AI products: lowering latency and cost while serving more tokens on the same hardware.

Key details: Published June 24, 2026 at 07:21 UTC, DFlash uses block diffusion drafting and KV injection, The reported benchmark is up to 15x higher throughput on Nvidia Blackwell, The technique targets LLM inference acceleration.

Why it matters: Inference throughput is where AI products either become affordable or stay constrained; large speedups on current accelerator platforms can matter as much as new model releases.