2026-05-07

DFlash

Original source

Reading Log X Post Historical Backfill AI Infra LLM Research

Imported from historical reading log.

  • Extracted main post via api.fxtwitter.com fallback and checked the linked repo z-lab/dflash.
  • Zhijian Liu pitches DFlash for Gemma 4 as an open-source speculative decoding path that can push native Gemma 4 MTP further, claiming up to 6x faster generation at the same quality.
  • Repo framing: DFlash: Block Diffusion for Flash Speculative Decoding, a lightweight block-diffusion draft model for speculative decoding, with support across Gemma, Qwen, Llama, GPT-OSS, MLX, vLLM, SGLang, and Transformers backends.
  • What seems notable is not just the speed claim, but that speculative-decoding acceleration is turning into a portable ecosystem layer with open models, backend integrations, and per-model draft variants.
  • Good angle: inference-speed competition is moving into open, pluggable speculative-decoding infrastructure.