Eagle-4 Released: Faster than DFlash2 and DSpark
EAGLE-4 has been released, achieving a 3.59x speedup over vanilla generation for the Qwen-27B model. It outperforms alternative acceleration techniques such as MTP, DSpark, and DFlash2. This release represents a significant performance milestone in LLM speculative decoding and acceleration, drastically reducing inference latency. Such advancements are crucial for making large language models more efficient and practical to run locally. According to initial benchmarks on the Qwen 3.8 27B model, EAGLE-4 delivers a 3.59x speedup, compared to 2.80x for DFlash 2, 2.57x for DSpark, and 2.22x for MTP. Users are currently anticipating integration into popular local serving frameworks like llama.cpp and vLLM.
## BACKGROUND
Speculative decoding is a technique used to speed up large language model inference by using a smaller, faster draft model to generate multiple candidate tokens in parallel, which are then verified simultaneously by the large target model. EAGLE is a popular speculative sampling framework that utilizes lightweight draft heads to extrapolate future tokens based on hidden states without losing generation quality.