Presented is the DFlash 2 drafter for accelerating the Qwen3.8-27B model

Developers introduced Qwen3.8-27B-DFlash2 — a specialized tool for accelerating text generation in the Qwen3.8-27B model. The solution is based on the DFlash 2 architecture, which acts as a 2B block-diffusion drafter.

The technology allows for inference acceleration of up to 3.43x when using 8 draft tokens at each generation step. During testing, the model showed a result of 236 points, spending 1 hour on task execution. Weights and configuration are available for download in the repository on the Hugging Face platform.

Using drafters (draft models) is an optimization method where a small model predicts a sequence of tokens, and the main model checks them in parallel. This allows significantly reducing the response time from LLM without losing the quality of generation characteristic of quantized versions.

The integration of DFlash 2 is aimed at developers using Qwen3.8-27B in high-load systems where low latency is critical. The tool allows for efficient scaling of work with the model on standard server equipment.