CMB-Qwen3-30B-A3B-DSpark-preview

CMB-Qwen3-30B-A3B-DSpark-preview

Model Overview

• Target Model: Qwen/Qwen3-30B-A3B
• Speculative Decoding Algorithm: DSpark
• Release Date: 08/17/2026
• Version: preview
• Model Developer: Architecture Management Team, Information Technology Department, China Merchants Bank (CMB)

Model Introduction

This model is implemented based on the DSpark speculative decoding algorithm, with Qwen/Qwen3-30B-A3B as the target model.

This model was trained and open-sourced by the Architecture Management Team of the Information Technology Department at China Merchants Bank. Built upon the Qwen3-30B-A3B architecture, it was trained using the CMB-Eagle-Mix-330k dataset regenerated in ` </think>

mode (withblock_size=8`).

This model should be used with the chat template of Qwen/Qwen3-30B-A3B and invoked via the /chat/completions API endpoint.

Deployment & Usage

Using vLLM

Note: When using vLLM, you must enable the dedicated configuration file config_vllm.json for this model. Before running the command below, please rename config_vllm.json in the model directory to config.json for it to take effect.

vllm serve Qwen/Qwen3-30B-A3B \
  --speculative-config '{
    "method": "dspark",
    "num_speculative_tokens": 7,
    "model": "<our-dspark-model-path>"
  }'

Evaluations

<h3>Evaluation Results (num_speculative_tokens = 7)</h3> <details> <summary><strong>Detailed Experimental Configuration</strong></summary> <br> <ul> <li><strong>Temperature:</strong> [0]</li> <li><strong>Top_p:</strong> [0.95]</li> <li><strong>Top_k:</strong> [20]</li> <li><strong>Repetitions:</strong> [1.0]</li> <li><strong>Hardware:</strong> [2×H800]</li> </ul> </details>