Skip to content
Trending storyDeveloping

Eagle Corp releases open-source EAGLE-4 speculative decoding drafter for LLM inference

2 articles2 sourcessince Oct 20Last article 1h ago ·

Overview

AISummary of 2 articles

Eagle Corp has released EAGLE-4, an open-source flexible drafter for speculative decoding, which speeds up LLM inference by having a smaller model propose tokens that the main model verifies.

Eagle Corp says one drafter adapts its strategy to batch size, hardware and workload while keeping the target model's output distribution. Its own benchmark reports 961.6 tokens/s for Qwen3.8-27B with EAGLE-4 on four NVIDIA B300 GPUs, using BF16, TP4 and one request at T=0, against 240.5 tokens/s for vanilla decoding. The same test reportedly gives 672.6 tokens/s for DFlash 2 and 618.6 for DSpark. Eagle Corp says the code can be run with vLLM and SGLang. These figures come from Eagle Corp's own test and have not been independently verified.

Written by AI from the articles below · updated Oct 20, 6:05 PM ET

Check the sources:

Latest developmentOct 20, 6:02 PM ET
Eagle Corp releases open-source EAGLE-4 drafter for faster LLM inference

Article timeline

The articles in this story. Times are ET.

Oct 20
  1. 🚨 AI News | TestingCatalogX
    Eagle Corp releases open-source EAGLE-4 drafter for faster LLM inference

    AIEagle Corp has released EAGLE-4, an open-source flexible drafter that adapts its speculative decoding to batch size, hardware and workload. In its own test, Qwen3.8-27B with EAGLE-4 reached 961.6 tokens/s on four NVIDIA B300 GPUs, against 240.5 tokens/s for vanilla decoding, 672.6 for DFlash 2 and 618.6 for DSpark.

    Image from @testingcatalog's post
  2. SantiagoX
    EAGLE-4 speculative decoding reports 961.6 tokens/s on Qwen3.8-27B

    AISantiago reposts EagleCorp's announcement of EAGLE-4, which it calls its fastest version yet. EagleCorp reports 961.6 tokens/s on Qwen3.8-27B using TokenSpeed with BF16, four NVIDIA B300 GPUs, TP4, one request, and T=0. The post explains that a smaller helper model guesses several upcoming tokens and the main model checks them together, and says the code is available to run with vLLM and SGLang.

Heat Trend

Not enough continuous observations to show a trend yet.