EAGLE-Corp/Qwen3.8-27B-EAGLE4

EAGLE-4 Draft Model for Qwen3.8-27B

One Flexible Drafter for All Batch Sizes and Hardware

This repository contains an EAGLE-4 draft model for the target model Qwen3.8-27B. The draft model is designed to be used with the EAGLE-4 implementation to accelerate inference through speculative decoding.

About EAGLE-4

EAGLE-4 (Extrapolation Algorithm for Greater Language-model Efficiency) is the latest generation of the EAGLE family for accelerating large language model inference.

EAGLE-4 introduces a single flexible drafter that dynamically adapts its speculative decoding strategy to different batch sizes, hardware platforms, and workloads. Its goal is to improve decoding performance while maintaining the target model's output distribution through speculative verification.

In speculative decoding, the draft model proposes candidate tokens that are checked by the target model. This checkpoint supplies the drafting component for Qwen3.8-27B; it is not a standalone text-generation model and should be used with the compatible EAGLE-4 code and target model.

Results

Throughput on the qualitative split of SPEED-Bench with Qwen3.8-27B at concurrency 1, using 4× NVIDIA B300 GPUs (TP=4). Maximum output length: 4,096 tokens. All methods use the same inference engine with MegaKernel disabled. Speedups are relative to Vanilla.

Temperature T=0

Method	Tokens/s	Speedup
EAGLE-4	863.8	3.59×
DFlash 2	672.6	2.80×
DSpark	618.6	2.57×
MTP	534.5	2.22×
Vanilla	240.5	1.00×

Temperature T=1

Method	Tokens/s	Speedup
EAGLE-4	608.6	2.59×
DFlash 2	560.3	2.39×
DSpark	519.4	2.21×
MTP	369.1	1.57×
Vanilla	234.9	1.00×

Resources

• EAGLE-4 code and setup: GitHub repository
• EAGLE-4 paper: PDF
• EAGLE website: eagle.ai
• EAGLE models: Hugging Face — EAGLE-Corp

License and Acknowledgements

See the EAGLE-4 repository's LICENSE and NOTICE for project licensing and acknowledgements. Check the license applicable to this checkpoint and the underlying Qwen3.8-27B weights before use. SPEED-Bench is by NVIDIA.

Citation

@misc{eagle4,
    author = {Eagle AI Team},
    title = {{EAGLE-4}: One Flexible Drafter for All Batch Sizes and Hardware},
    year = {2026},
    note = {Preprint}
}

@inproceedings{li2024eagle,
    author = {Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang},
    title = {{EAGLE}: Speculative Sampling Requires Rethinking Feature Uncertainty},
    booktitle = {International Conference on Machine Learning},
    year = {2024}
}

@inproceedings{li2024eagle2,
    author = {Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang},
    title = {{EAGLE-2}: Faster Inference of Language Models with Dynamic Draft Trees},
    booktitle = {Empirical Methods in Natural Language Processing},
    year = {2024}
}

@inproceedings{li2025eagle3,
    author = {Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang},
    title = {{EAGLE-3}: Scaling up Inference Acceleration of Large Language Models via Training-Time Test},
    booktitle = {Annual Conference on Neural Information Processing Systems},
    year = {2025}
}