EAGLE

| Paper (EAGLE) | Paper (EAGLE-2) | Paper (EAGLE-3) | Blog |

EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency) is a new baseline for fast decoding of Large Language Models (LLMs) with provable performance maintenance. This approach involves extrapolating the second-top-layer contextual feature vectors of LLMs, enabling a significant boost in generation efficiency.

• EAGLE is:
  • certified by the third-party evaluation as the fastest speculative method so far.
  • achieving 2x speedup on gpt-fast.
  • 3x faster than vanilla decoding (13B).
  • 2x faster than Lookahead (13B).
  • 1.6x faster than Medusa (13B).
  • provably maintaining the consistency with vanilla decoding in the distribution of generated texts.
  • trainable (within 1-2 days) and testable on 8x RTX 3090 GPUs. So even the GPU poor can afford it.
  • combinable with other parallelled techniques such as vLLM, DeepSpeed, Mamba, FlashAttention, quantization, and hardware optimization.

EAGLE-2 uses the confidence scores from the draft model to approximate acceptance rates, dynamically adjusting the draft tree structure, which further enhances performance.

• EAGLE-2 is:
• 4x faster than vanilla decoding (13B).
• 1.4x faster than EAGLE-1 (13B).

EAGLE-3 removes the feature prediction constraint in EAGLE and simulates this process during training using training-time testing. Considering that top-layer features are limited to next-token prediction, EAGLE-3 replaces them with a fusion of low-, mid-, and high-level semantic features. EAGLE-3 further improves generation speed while ensuring lossless performance.

• EAGLE-3 is:
• 5.6 faster than vanilla decoding (13B).
• 1.8x faster than EAGLE-1 (13B).

Inference is conducted on 2x RTX 3090 GPUs at fp16 precision using the Vicuna 13B model.

Support

EAGLE has been merged in the following mainstream LLM serving frameworks (listed in alphabetical order).

• AMD ROCm
• AngelSlim
• AWS NeuronX Distributed Core
• CPM.cu
• Intel® Extension for Transformers
• Intel® LLM Library for PyTorch
• MLC-LLM
• NVIDIA NeMo Framework
• NVIDIA TensorRT-LLM
• NVIDIA TensorRT Model Optimizer
• PaddleNLP
• SGLang
• SpecForge
• speculators
• vLLM

Update

2025.9.18: EAGLE-3 is accepted to NeurIPS'25.

2025.7.23: We strongly recommend using SpecForge for out-of-the-box training of EAGLE-3 with SGLang.

2025.3.19: EAGLE-3 is released.

2024.8.8: We now support Qwen-2.

2024.6.27: EAGLE-2 is released.

2024.2.25: EAGLE is certified by the third-party evaluation as the fastest speculative method.

2024.1.17: We now support Mixtral-8x7B-Instruct.

2023.12.8: EAGLE v1.0 is released.

Todo

• [x] Support non-greedy inference (provably maintaining text distribution).
• [x] Support more LLMs such as Mixtral 8x7B.
• [x] Support LLaMA-3.
• [x] Support Qwen-2.
• [x] Support vLLM (please check vLLM's implementation).
• [x] EAGLE-3.
• [x] Training code of EAGLE-3.
• [x] Support LLaMA-4.
• [ ] Support official EAGLE-3 for Qwen-3.
• [ ] EAGLE-4.

The default main branch is the implementation of EAGLE-3 and EAGLE-2. For using EAGLE-1, please switch to the v1 branch.

Contents

• Setup & Installation
• EAGLE-3 Weights
• EAGLE Weights
• Inference
• With UI
• With Code
• Train
• Generate Train Data
• Train the Auto-regression Head
• Inference on custom models
• Evaluation

Setup & Installation

git clone https://github.com/SafeAILab/EAGLE.git
cd EAGLE
python -m venv ~/venvs/ea_env
source ~/venvs/ea_env/bin/activate
pip install -r requirements.txt

EAGLE-3 Weights

Note: This repository recognizes only official EAGLE-3 checkpoints. Performance of unofficial checkpoints may vary. If you want to compare with EAGLE-3, please compare with official checkpoints and official draft tree setups.

EAGLE-3 Models on Hugging Face

Base Model	EAGLE-3 Model(s)	Official
Vicuna-13B v1.3

lmsys/vicuna-13b-v1.3 | yuhuili/EAGLE3-Vicuna1.3-13B | Yes | | LLaMA-3.1-8B-Instruct

meta-llama/Llama-3.1-8B-Instruct | yuhuili/EAGLE3-LLaMA3.1-Instruct-8B | Yes | | LLaMA-3.3-70B-Instruct

meta-llama/Llama-3.3-70B-Instruct | yuhuili/EAGLE3-LLaMA3.3-Instruct-70B | Yes | | DeepSeek-R1-Distill-LLaMA-8B

deepseek-ai/DeepSeek-R1-Distill-Llama-8B | yuhuili/EAGLE3-DeepSeek-R1-Distill-LLaMA-8B | Yes | | LLaMA-4-Scout-17B-16E-Instruct

meta-llama/Llama-4-Scout-17B-16E-Instruct | lmsys/sglang-EAGLE3-Llama-4-Scout-17B-16E-Instruct-v1 | No | | LLaMA-4-Maverick-17B-128E-Instruct

meta-llama/Llama-4-Maverick-17B-128E-Instruct | lmsys/sglang-EAGLE3-Llama-4-Maverick-17B-128E-Instruct-v1

nvidia/Llama-4-Maverick-17B-128E-Eagle3 | No | | Qwen3-1.7B

Qwen/Qwen3-1.7B | AngelSlim/Qwen3-1.7B_eagle3 | No | | Qwen3-4B

Qwen/Qwen3-4B | AngelSlim/Qwen3-4B_eagle3 | No | | Qwen3-8B

Qwen/Qwen3-8B | Tengyunw/qwen3_8b_eagle3

AngelSlim/Qwen3-8B_eagle3

Zjcxy-SmartAI/Eagle3-Qwen3-8B-zh | No | | Qwen3-14B

Qwen/Qwen3-14B | AngelSlim/Qwen3-14B_eagle3 | No | | Qwen3-30B-A3B

Qwen/Qwen3-30B-A3B | Tengyunw/qwen3_30b_moe_eagle3

AngelSlim/Qwen3-a3B_eagle3 | No | | Qwen3-32B

Qwen/Qwen3-32B | AngelSlim/Qwen3-32B_eagle3

Zjcxy-SmartAI/Eagle3-Qwen3-32B-zh | No | | Qwen3-235B-A22B

Qwen/Qwen3-235B-A22B | nvidia/Qwen3-235B-A22B-Eagle3

lmsys/Qwen3-235B-A22B-EAGLE3 | No | | MiniCPM4-8B

openbmb/MiniCPM4-8B | linglingdan/Eagle3_for_MiniCPM4 | No | | OLMoE-1B-7B-Instruct

allenai/OLMoE-1B-7B-0125-Instruct | wantsleep/OLMoE_1B_7B_Eagle3 | No | | granite-3.1-1b-a400m-instruct

ibm-granite/granite-3.1-1b-a400m-instruct | wantsleep/granite-3.1-1b-a400m-EAGLE3 | No | | GPT-OSS-120B

openai/gpt-oss-120b | lmsys/EAGLE3-gpt-oss-120b-bf16

nvidia/gpt-oss-120b-Eagle3 | No | | GLM-4.7-Flash

zai-org/GLM-4.7-Flash | thoughtworks/GLM-4.7-Flash-Eagle3 | No |

EAGLE Weights

Note: The current code defaults to using EAGLE-3. If you want to use EAGLE weights, please specify use_eagle3=False in EaModel.from_pretrained.

Note: When Qwen2 is the target model, please use bf16 precision instead of fp16 to avoid numerical overflow. The training dataset for the draft model of Qwen2 is ShareGPT, which has removed non-English data. Therefore, if you want to use it on non-English data such as Chinese, please train with the corresponding data.

EAGLE Models on Hugging Face

Base Model	EAGLE Model	# EAGLE Parameters	Official
Vicuna-7B v1.3	yuhuili/EAGLE-Vicuna-7B-v1.3	0.24B	Yes
Vicuna-13B v1.3	yuhuili/EAGLE-Vicuna-13B-v1.3	0.37B	Yes
Vicuna-33B v1.3	yuhuili/EAGLE-Vicuna-33B-v1.3	0.56B	Yes
LLaMA2-Chat 7B	yuhuili/EAGLE-llama2-chat-7B	0.24B	Yes
LLaMA2-Chat 13B	yuhuili/EAGLE-llama2-chat-13B	0.37B	Yes
LLaMA2-Chat 70B	yuhuili/EAGLE-llama2-chat-70B	0.99B	Yes
Mixtral-8x7B-Instruct v0.1	yuhuili/EAGLE-mixtral-instruct-8x7B	0.28B	Yes
LLaMA3-Instruct 8B	yuhuili/EAGLE-LLaMA3-Instruct-8B	0.25B	Yes
LLaMA3-Instruct 70B	yuhuili/EAGLE-LLaMA3-Instruct-70B	0.99B	Yes
Qwen2-7B-Instruct	yuhuili/EAGLE-Qwen2-7B-Instruct	0.26B	Yes
Qwen2-72B-Instruct	yuhuili/EAGLE-Qwen2-72B-Instruct	1.05B	Yes
LLaMA3.1-Instruct 8B	[yuhuili/EAGLE-LLaMA3.1-Instruct-8B](https://huggingface.co/yuhuili/EAGLE-LLaMA3.1-Instru