EAGLE-4 Speculative Decoding Open-Sourced

Eagle AI Corp has officially released its fourth-generation speculative decoding method, EAGLE-4, along with the official inference code as open source, pitching it as "One Flexible Drafter for All Batch Sizes and Hardware." Its official positioning, open-source repository, and multiple benchmark results have been cross-confirmed by several sources.

Confirmed

The official inference code is now open source on GitHub (m1, m5); the repo is published under the name eagleharness (m2) and serves speculative decoding drafts to SGLang, vLLM, and TokenSpeed via engine plugins (m2); the repo had 162 stars at launch (m5)
Officially, it reached 961.6 tokens/s on Qwen3.8-27B (BF16, 4×B300, TP4, single request, T=0) paired with TokenSpeed (m3)
Speedup comparison on Qwen3.8-27B: EAGLE-4 achieves 3.59×, beating DFlash 2 (2.80×), DSpark (2.57×), and MTP (2.22×), making it the fastest speculative decoding method on this model to date (m1)
A drafter for Qwen3.8-27B was released alongside (m5)
LightSeek published a technical deep dive on EAGLE-4 training and serving: the drafter is trained with TorchSpec and full inference support is provided by TokenSpeed; the speculative process is lossless; agentic coding workloads see speedups of up to 5.21× (m4)

Why it matters

Speculative decoding is a key path to cutting LLM inference cost and latency. EAGLE-4 covers varying batch sizes and hardware with a single drafter while keeping outputs lossless, directly serving deployment on mainstream inference frameworks like SGLang and vLLM
Official data and third-party results (LightSeek's write-up and TKGaming11's benchmark recounts) corroborate each other, showing EAGLE-4 genuinely leads competing methods DFlash 2 and DSpark on Qwen3.8-27B
Both the training (TorchSpec) and serving (TokenSpeed) pipelines are publicly documented, making it easy for the community to reproduce and build on

2026-10-21 ~ 2026-10-21 · 5 related posts

Primary sources

EAGLE-4 hits 3.59x decoding speedup on Qwen3.8-27B, beating DFlash 2 and DSpark — TKGaming_11 · 2026-10-21 07:04 (UTC+08:00)
EAGLE-4 speculative decoding goes open source with SGLang and vLLM plugins — testingcatalog · 2026-10-21 06:02 (UTC+08:00)
EAGLE-4 deep dive: up to 5.21× agentic coding speedup with lossless speculative decoding — zhyncs42 · 2026-10-21 05:49 (UTC+08:00)

EAGLE-4 speculative decoding hits 961.6 tokens/s on Qwen3.8-27B, code open-sourced — zhyncs42 · 2026-10-21
[source] EAGLE-4 deep dive: up to 5.21× agentic coding speedup with lossless speculative decoding — zhyncs42 · 2026-10-21
[source] EAGLE-4 speculative decoding goes open source with SGLang and vLLM plugins — testingcatalog · 2026-10-21
[source] EAGLE-4 hits 3.59x decoding speedup on Qwen3.8-27B, beating DFlash 2 and DSpark — TKGaming_11 · 2026-10-21
EAGLE-4 speculative decoding inference code goes open source on GitHub — rohanpaul_ai · 2026-10-21