CoolFace
Modelpublic

CMBTech/CMB-Qwen3-32B-Eagle3

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
0likes32downloads
Model Card

CMB-Qwen3-32B-Eagle3

Model Overview

  • —Verifier: Qwen/Qwen3-32B
  • —Speculative Decoding Algorithm: EAGLE-3
  • —Model Architecture: LlamaForCausalLMEagle3
  • —Release Date: 03/31/2026
  • —Version: 1.0
  • —Model Developers: Architecture Management Team, Information Technology Department, China Merchants Bank (CMB)

This is a speculator model designed for use with Qwen/Qwen3-32B, based on the EAGLE-3 speculative decoding algorithm. It was trained by Architecture Management Team, Information Technology Department, China Merchants Bank (CMB). The training data is a combination of Chinese and English open-source datasets, with data synthesis and cleaning performed on the original datasets. The following are the sources of the original training data:

  1. 1.swift/Chinese-Qwen3-235B-Thinking-2507-Distill-data-110k-SFT
  2. 2.a-m-team/AM-Qwen3-Distilled

This model should be used with the Qwen/Qwen3-32B chat template, specifically through the /chat/completions endpoint.

Use with sglang

bash
python3 -m sglang.launch_server \
  --model Qwen/Qwen3-32B \
  --speculative-algorithm EAGLE3 \
  --speculative-draft-model-path <our-model-path> \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 3

Use with vLLM

It is important to note that when using vLLM, you need to enable the model configuration file for vLLM, which named config_vllm.json. Before entering the following commands, you need to rename the file from config_vllm.json to config.json to make it effective.

bash
vllm serve Qwen/Qwen3-32B \
  -tp 2 \
  --speculative-config '{
    "model": <our-model-path>,
    "num_speculative_tokens": 3,
    "method": "eagle3"
  }'

Evaluations

<h3>Use cases</h3> <table> <thead> <tr> <th>Use Case</th> <th>Dataset</th> <th>Samples</th> </tr> </thead> <tbody> <tr> <td>Coding</td> <td>HumanEval</td> <td>168</td> </tr> <tr> <td>Math Reasoning</td> <td>gsm8k</td> <td>80</td> </tr> <tr> <td>English Text Summarization</td> <td>CNN/Daily Mail</td> <td>80</td> </tr> <tr> <td>Chinese Mixed</td> <td>HC3-Chinese</td> <td>120</td> </tr> <tr> <td>Chinese Finance</td> <td>HC3-Chinese</td> <td>80</td> </tr> <tr> <td>Chinese Text Summarization</td> <td>LCSTS</td> <td>80</td> </tr> </tbody> </table>

<h3>Acceptance lengths</h3> <table> <thead> <tr> <th>Use Case</th> <th>drafttokensnum=3</th> </tr> </thead> <tbody> <tr> <td>Coding</td> <td>2.11</td> </tr> <tr> <td>Math Reasoning</td> <td>2.60</td> </tr> <tr> <td>English Text Summarization</td> <td>1.95</td> </tr> <tr> <td>Chinese Mixed</td> <td>2.08</td> </tr> <tr> <td>Chinese Finance</td> <td>2.15</td> </tr> <tr> <td>Chinese Text Summarization</td> <td>2.05</td> </tr> </tbody> </table>

<details> <summary>Details</summary> <strong>Configuration</strong>

  • —temperature: 0.6
  • —top_p: 0.95
  • —top_k: 20
  • —repetitions: 3
  • —hardware: 2xH800

</details>