What changed

Tencent added Qwen3.8-27B support to DSpine and released an Apache-2.0 draft checkpoint for speculative decoding. The documented SGLang path supports one GPU, as many as eight concurrent requests and 131,072-token target context with workload-specific verification-budget tuning.

Why it matters

The release creates a downloadable acceleration path for a current 27B model rather than only a paper design. Teams can now test whether a small parallel drafter improves their own latency and throughput without changing target-model outputs.

What to do

Pin the DSpine repository, SGLang 0.5.22 and the exact Qwen3.8-27B draft checkpoint; tune the SPS table on representative prompts, then compare acceptance length, tokens per second, memory, tail latency and output equality against the target model without speculation.

Caveat

Tencent publishes benchmark results for Qwen3-8B, not the new Qwen3.8-27B path. The 27B checkpoint's model card is sparse, the documented command trusts remote code, and actual gains depend on GPU, concurrency, prompt length and tuning.