---
title: "Tencent adds DSpine for Qwen3.8-27B, targeting faster self-hosted inference"
description: "Tencent adds DSpine support for Qwen3.8-27B. Compare five developer-reported throughput results and practical checks for self-hosted inference."
url: "https://promptszone.com/tencent-dspine-qwen3-8-27b-inference-throughput"
published: "2026-11-08T19:00:05Z"
categories: ["New Models"]
---

# Tencent adds DSpine for Qwen3.8-27B, targeting faster self-hosted inference

Tencent's new draft checkpoint accelerates Qwen3.8-27B inference in developer replay tests. Compare the results and what to measure before deployment.

![Tencent-reported xhigh replay throughput comparing autoregressive decoding, DFlash2 and DSpine across five workloads, in output tokens per second.](https://promptszone.com/assets/uploads/213f19bf1863da9fd0379a1af71af761.png)

**Summary:**

DSpine adds a speculative-decoding option for Qwen3.8-27B. Five replay comparisons show reported throughput gains, with configuration differences that matter for real deployments.

Tencent has added a DSpine draft checkpoint for Qwen3.8-27B, giving self-hosting teams another way to accelerate token generation. The November repository update provides inference support and links to the weights. Its practical appeal is throughput, rather than a new general-purpose assistant.

## What changes in a serving stack

DSpine proposes multiple tokens in parallel for a target model to verify. Tencent's repository describes causal information passing through the drafter's layers, rather than drafting every token sequentially. The Qwen3.8-27B setup uses 16-token draft blocks and an 8K sliding window.

The [official repository](https://github.com/Tencent/DSpine) supplies a single-GPU serving example for up to eight concurrent requests, using SGLang and additional verification optimizations. It also recommends tuning verification budgets for the actual GPU and workload. A working command is a starting configuration, not a capacity guarantee for every machine.

## Read the speed comparison as a deployment experiment

Tencent's [model card](https://huggingface.co/tencent/Qwen3.8-27B-DSpine-b16) compares autoregressive decoding, DFlash2 and DSpine. The table below reproduces its xhigh agent-replay throughput in output tokens per second; higher is better, and each row's highest value is bold.

| Replay workload | Autoregressive | DFlash2 | DSpine |
| --- | --- | --- | --- |
| WorkBuddy Code | 193.00 | 550.40 | **613.55** |
| WorkBuddy Web | 252.26 | 734.04 | **840.14** |
| WorkBuddy Office | 215.05 | 662.34 | **763.83** |
| SWE-bench Pro | 248.17 | 711.60 | **791.30** |
| τ²-bench | 340.39 | 871.98 | **978.39** |

These are developer-reported measurements with SGLang 0.5.22, BF16 and concurrency eight. Each subset uses 20 conversations with recorded tool-decision prefixes replayed and prefix caching enabled. DSpine uses 16-token blocks; DFlash2 uses eight. The comparison therefore evaluates these configured systems, not an isolated change with every setting identical.

A SWE-bench-labelled row here measures generation speed during replay. It is not the percentage of software issues an agent resolved. Nor does aggregate tokens per second tell a team how long its first user will wait for an answer.

## Three checks before adopting it

- **Match the target.** This checkpoint is for Qwen3.8-27B. Do not assume an accelerator trained for one target transfers to another version.
- **Measure the experience.** Compare time to first token, total answer latency and memory at your own concurrency. Include short requests, long contexts and realistic tool traffic.
- **Keep a rollback baseline.** Retain the existing decoding path and compare outputs and operational failures before routing production traffic to the new setup.

The model card identifies an Apache 2.0 license for the draft checkpoint. Teams still need to account for the target model and the rest of their serving software separately.

## Sources and image credit

Checked November 8, 2026: [Tencent DSpine repository and November update](https://github.com/Tencent/DSpine); [official checkpoint, methodology and results](https://huggingface.co/tencent/Qwen3.8-27B-DSpine-b16). The chart is an original visualization of Tencent's reported data. The unaltered official Tencent provider logo is used as the fallback brand mark from its [editorial-use media library](https://www.tencent.com/newsroom/media-resources/logos/); it is not a distinct DSpine logo.