---
title: SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan
canonical_url: "https://www.modelscope.cn/models/lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan"
md_url: "https://www.modelscope.cn/models/lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan.md"
repository: lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan
last_updated: 2025-12-28
license: mit
model_type:
  - llama
architectures:
  - LlamaForCausalLMEagle3
parameters: 592.6M
tensor_type:
  - BF16
  - BOOL
  - I64
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
inference_backends:
  - "deploy_task text"
  - "sglang 0.5.2"
downloads: 175
stars: 0
---

# SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan

> SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan - lmsys 在 ModelScope 开源的模型。EAGLE3 For Qwen/Qwen3-235B-A22B-Instruct-2507

lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan 是 ModelScope 魔搭社区上的 592.6M 参数机器学习模型，采用 mit 许可，可用 deploy_task text、sglang 0.5.2 部署。

- **Repository**: lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan
- **License**: mit
- **Parameters**: 592.6M
- **Inference backends**: deploy_task text, sglang 0.5.2
- **Downloads**: 175
- **Stars**: 0
- **Last updated**: 2025-12-28

Source: https://www.modelscope.cn/models/lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan

---

![meituan_team](https://cdn-uploads.huggingface.co/production/uploads/63993d721fad4d6eb265d999/vCqSrLcr-6Z1XRBnBR0sB.png)

# EAGLE3 For Qwen/Qwen3-235B-A22B-Instruct-2507


## About

**SpecBundle** is an open-source initiative, jointly driven by the community and industry, to democratize speculative decoding by providing high-performance speculative decoding draft weights for mainstream open-source models.

This checkpoint was trained by the [**Meituan**](https://huggingface.co/meituan) Team and released as the **phase 1 of SpecBundle release**. We regenerated the responses in the [mlabonne/open-perfectblend](https://huggingface.co/datasets/mlabonne/open-perfectblend) and trained the model on 1.4M data samples for 2 epochs. This checkpoint was trained using the [**SpecForge**](https://github.com/sgl-project/SpecForge) framework.



## Usage

You can use this checkpoint with the command below.

```bash
export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1

python3 -m sglang.launch_server \
    --model Qwen/Qwen3-235B-A22B-Instruct-2507 \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-perfect-blend-regenerated \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --tp 8
```

# Performance

This checkpoint exhibits superior performance on various benchmarks.

<table>
  <tr>
    <th>Throughput</th>
    <th>Acceptance Length</th>
  </tr>
  <tr>
    <td><img src="https://cdn-uploads.huggingface.co/production/uploads/63993d721fad4d6eb265d999/VT7KzWJDuS62ETMi5eNba.png" /></td>
    <td><img src="https://cdn-uploads.huggingface.co/production/uploads/63993d721fad4d6eb265d999/x4Rbh7sqBy9iBM-Kyrslw.png" /></td>
  </tr>
</table>


You can reproduce the performance with the command below:

```bash
# clone specforge
git clone https://github.com/sgl-project/SpecForge.git
cd SpecForge/benchmarks

# run benchmarks
python bench_eagle3.py \
        --model Qwen/Qwen3-235B-A22B-Instruct-2507 \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-perfect-blend-regenerated \
        --port 30003 \
        --config-list 1,3,1,4 1,5,1,6 1,5,3,6 1,7,1,8 1,7,4,8 \
        --benchmark-list gsm8k math500 mtbench humaneval livecodebench financeqa gpqa  \
        --dtype bfloat16 \
        --tp 8 \
        --name qwen3-235b-a22b-spec-bundle
```

## Acknowledgement

We sincerely appreciate the collective efforts from both the developers in the open-source community and our industrial partners, especially Ant Group AQ Team, Meituan, Nex-AGI (Qiji Zhifeng), EigenAI for their invaluable contributions to the release of SpecBundle Phase 1.