Eagle3

#106 in Open-Source LLMs

DeepSeek · v3 · 9× · last seen Jul 01, 2026

7

Momentum

Eagle3 is a Speculative Decoding algorithm for faster LLM Inference, included as one of three Draft-Model methods (alongside DSpark and DFlash) in the official DeepSeek repository "DeepSpec". A lightweight Draft model predicts multiple Tokens in advance, which the large target model verifies in parallel in a verification step, accelerating generation without changing output quality. In DeepSeek's own Benchmarks on Qwen3 models (4B/8B/14B), Eagle3 is outperformed by DeepSeek's new DSpark method by 26.7–30.9% in accepted Token length. Eagle3 itself is based on the academic EAGLE-3 method (SafeAILab/NeurIPS'25), whose training and evaluation code has been integrated into DeepSpec.

Sources (9)