Qwen3-8B-DSpine-b16: 4GB+ VRAM

Qwen3-8B-DSpine-b16: 4GB+ VRAM

At a Glance

Item Value
Repository tencent/Qwen3-8B-DSpine-b16
Publisher guide Tencent: models and licenses
Published 2026-10-16
License apache-2.0
Formats safetensors
Paper arXiv:2609.36173
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Tencent has released “DSpine", a parallel draft model designed for speculative decoding.

In conventional parallel draft methods, each position selects tokens independently, leading to inconsistent continuation content and a shorter accepted prefix. DSpine uses gated adjacent injection at each layer of the backbone to write the prediction features of the immediate preceding position into each position. This makes it possible to unfold a chain of causal conditioning along the depth of the network while updating all positions in parallel.

According to the analysis in the paper, existing methods already form recoverable predictions in shallow layers, but accurate neighbor information is not being propagated to subsequent tokens. DSpine utilizes a unified transfer space constructed from the target model’s output embeddings, integrating layer-by-layer injection and neighbor-prediction-based decoding to promote the formation of prediction features in shallow layers. For more details, refer to the paper arXiv:2609.36173.

Specifications

  • Architecture: DFlashDraftModel
  • Model Type: qwen3
  • Parameters: 8B
  • Context Length: 8192

Performance

The comparison results of the acceptance length on Qwen3-8B published in the model card are as follows. Note that all comparison models were retrained using the same Open-PerfectBlend dataset with identical configurations.

→ Scroll horizontally to see all columns

Temp. Method GSM8K MATH-500 HumanEval MBPP LCB MT-Bench Arena-Hard Avg.
T = 0 DFlash† 6.32 7.99 6.56 6.29 6.11 4.26 3.91 5.92
Domino† 9.68 9.40 7.23 6.89 6.40 5.12 4.65 7.05
DSpark 10.00 9.88 7.78 7.35 6.81 5.39 4.87 7.44
DFlash2 10.06 10.04 7.99 7.33 6.96 5.40 5.06 7.55
DSpine 10.55 10.30 8.27 7.66 7.25 5.63 5.13 7.83
T = 1 DFlash† 5.75 6.51 5.52 5.23 4.53 3.77 3.25 4.94
Domino† 8.01 7.13 5.89 5.56 4.59 4.36 3.61 5.59
DSpark 9.33 8.64 6.91 6.66 5.78 4.99 4.25 6.65
DFlash2 9.07 8.42 6.99 6.43 5.67 4.90 4.25 6.53
DSpine 9.83 8.99 7.39 6.97 6.11 5.09 4.49 6.98

According to the measurement results in the model card, DSpine records the highest acceptance length across all benchmark items in both T=0 and T=1 temperature settings. It outperforms the comparison models in mathematics capabilities measured by GSM8K and MATH-500, coding capabilities measured by HumanEval and MBPP, and Arena-Hard, an indicator of conversation quality. In particular, it can be seen that it consistently achieves higher figures compared to existing methods such as DFlash2 and DSpark. Additionally, according to the abstract of the paper, serving tests using SGLang report an average of 23.3% higher throughput compared to DFlash.

Strengths and Use Cases

DSpine is a draft model aimed at improving the efficiency of speculative decoding. Therefore, it is specialized for the purpose of accelerating the inference speed of target large language models.

From the tags in the model card and the description in the paper, it excels at maximizing the “acceptance length" in the inference of models like the Qwen3 series. This is expected to bring out high performance while maintaining inference efficiency across a wide range of tasks such as mathematics (GSM8K, MATH-500), coding (HumanEval, MBPP), and dialogue tasks (Arena-Hard).

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 1.2B parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 2.3GB 2.8GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

How to Get It

  • Distribution Format: safetensors
  • Installation Command:
git clone https://github.com/tencent/dspine.git
cd dspine
pip install -e.
  • Supported Engine: SGLang

Related Articles

What to Read Next

Sources