Qwen3.8-27B-DSpine-b16

About

A five-layer draft model for speculative decoding with Qwen3.8-27B, using a block size of 16 and an 8,192-token sliding window. It generates draft tokens in parallel, so the target model can verify several tokens per step. In the card's SGLang tests at concurrency 8, it reached longer accepted draft lengths than the official DFlash2 baseline and agent throughput of roughly 2.9x to 3.8x over autoregressive decoding.