Qwen3.8-27B-DSpine-b16: Tencent's parallel draft model for faster Qwen3.8 serving
Tencent's 2B-parameter draft model for Qwen3.8-27B that guesses 16 tokens at once so the big model verifies them in a single pass, reaching up to 4x throughput over autoregressive decoding.
Why it matters
Speculative decoding with causal conditioning injected layer by layer: acceptance length beats DFlash2 on every benchmark, and the weights are Apache 2.0 with an SGLang plugin ready to run.
What you could build with it
A team serving Qwen3.8-27B behind an API could plug this draft model into SGLang and cut per-token latency for agent workloads, serving more concurrent users on the same GPUs without changing the main model.
Does it hold up?
Too early to judge: the 27B weights landed November 9, the numbers are vendor-measured on a specific SGLang setup, and real-world speedups depend on each deployer's workload.
Built with Qwen3.8-27B-DSpine-b16
- What Is DSpine? Tencent's Speculative Decoding Draft Modelarticle · Plain-language explainer: a 5-layer, 2B-parameter helper for Qwen3.8-27B, tied to SGLang 0.5.22, with the GSM8K acceptance-length gains over DFlash2.
- DSpine: Depth-Wise Causal Injection for Parallel Speculative Decodingarticle · Overview of the method and results: up to 4x throughput on SGLang, now serving Tencent WorkBuddy inference workloads, with paper and code links.
- Tencent launches Qwen3.8-27B-DSpine-b16 draft model on Hugging Facearticle · Launch notice: five-layer draft model, block size 16, 8,192-token sliding window, Apache 2.0, with SGLang serving instructions.
- tencent/dspinegithub · 252 ★ · Official DSpine code and SGLang plugin: clone, install, and serve Qwen3.8-27B with the draft model via --speculative-algorithm DSPINE.
- sgl-project/sglanggithub · 38,900 ★ · The serving runtime DSpine ships as a plugin for; SGLang 0.5.22 is the measured configuration behind the reported throughput gains.
Learn more
First spotted on hf: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Google AI Edge Gallery: offline speech recognition and synthesis on-deviceGoogle added fully offline speech recognition and text-to-speech to AI Edge Gallery, with ML Kit APIs on Android and…models · JEV 0.73Soup: fine-tune an 8B LLM on a 4 GB laptop GPU from one YAML fileAn open-source CLI that fine-tunes LLMs from a single YAML file, with layer streaming that trains an 8B model on a 4…models · JEV 0.73Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.