Descriptions:
ZLab, the UC San Diego research team that invented DFlash speculative decoding, has released the first official drafter model paired with Google’s Gemma 4 26B — a mixture-of-experts architecture that activates only 4 billion of its 26 billion parameters per token, delivering the speed of a small model with the reasoning depth of a much larger one. Unlike earlier community-built drafters, this is a first-party release from the original inventors, making it a significant milestone for the DFlash ecosystem.
In this hands-on walkthrough, the presenter demonstrates a full local deployment on an Nvidia H100 with 80GB of VRAM using vLLM. The setup runs the two models in tandem: Gemma 4 26B as the primary model and the ZLab DFlash drafter proposing 15 tokens per step in parallel for bulk verification in a single forward pass. The vLLM command uses Triton as the attention backend for the main model and FlashAttention for the smaller drafter, with a 32K max batched token window.
Viewers learn how block diffusion differs from standard sequential speculative decoding, how to configure the vLLM serve command for dual-model inference, and what real generation output looks like on a complex HTML animation prompt. The model is gated on Hugging Face but released under Apache 2.0, and the video covers the access approval steps alongside GPU rental options for those without local hardware.
📺 Source: Fahd Mirza · Published May 07, 2026
🏷️ Format: Tutorial Demo
