DFlash: Block Diffusion for Flash Speculative Decoding - Z Lab

DFlash: Block Diffusion for Flash Speculative Decoding - Z Lab

4.0

DFlash is a cutting-edge technology by Z Lab that accelerates large language model inference through block diffusion and speculative decoding, delivering up to 3x speedup without compromising quality.

Last Updated Nov 10, 2026
DFlash: Block Diffusion for Flash Speculative Decoding - Z Lab screenshot

About

DFlash: Block Diffusion for Flash Speculative Decoding is a cutting-edge technology developed by Z Lab, designed to tackle the efficiency bottlenecks in large language model inference. Traditional autoregressive decoding methods deliver high accuracy, but generation speed is constrained by token-by-token computation, making it difficult to meet the demands of real-time applications. By combining a block diffusion mechanism with speculative decoding, DFlash significantly boosts inference throughput without compromising generation quality. This technology is especially well-suited for fast-response dialogue systems, content generation tools, and edge deployment scenarios, offering developers an efficient and scalable solution.

Key Features

Block Diffusion Speculative Decoding: Uses a diffusion model to generate multiple candidate token blocks in parallel, paired with a lightweight verifier for rapid screening, substantially reducing the number of decoding steps.
Low-Latency Inference: Leverages precomputation and cache optimization to cut single-inference latency down to 30%-50% of traditional methods.
Seamless Integration: Supports mainstream deep learning frameworks such as PyTorch and TensorFlow, and provides a clean API interface for easy integration into existing workflows.
Dynamic Resource Tuning: Automatically adjusts block size and parallelism based on hardware configuration to balance speed and memory usage.

Pricing & Fees

  • Pricing: Not publicly listed on the pages we reviewed; check the official site for current plans.
  • Open code and models: DFlash's paper, code, and pretrained draft models (e.g. z-lab/Qwen3-8B-DFlash-b16, z-lab/Qwen3-Coder-30B-A3B-DFlash) are publicly released for download and self-hosting.
  • Framework availability: Supports SGLang for production serving and Hugging Face Transformers for quick experimentation; vLLM integration is in progress.
  • Performance claim: Reports up to 6× lossless acceleration on Qwen3-8B, roughly 2.5× faster than EAGLE-3, on the same hardware.

Who It's For

AI researchers and engineers looking to explore new decoding paradigms and optimize large model inference performance; product developers who need to integrate real-time AI capabilities such as chatbots or code completion into their applications; edge computing practitioners deploying large models on resource-constrained devices while prioritizing low power consumption and fast response; and enterprise technology decision-makers evaluating options to reduce LLM serving costs and improve user experience.