Tencent says DSpine speeds Qwen3.8-27B inference in agent tests
The five-layer speculative-decoding drafter delivered 2.71x to 3.84x the autoregressive baseline in Tencent's agent throughput tests.
By RuntimeWire Staff · Published
Primary source: Aligned News — AI Intelligence
Why it matters
Inference costs affect the practicality of open-weight models when agent workloads produce long, tool-heavy responses. A downloadable drafter that raises accepted-token length can cut latency and serving cost without replacing the target model.
Tencent lists a five-layer DSpine draft checkpoint for Qwen3.8-27B on Hugging Face. The GitHub repository records Qwen3.8-27B inference support only at month level, as "[2026-11] Added support for Qwen3.8-27B inference," and gives no day for that entry. The repository roadmap separately lists a Qwen3.8-27B draft weight release, so the month-level support entry should not be read as a dated record of when the checkpoint was first uploaded.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.