EAGLE-4 Arrives With Adaptive Speculative Decoding

On October 20, 2026, Eagle AI Corp published EAGLE-4's first public inference-code release. The v0.1.0 release packages an adaptive speculative-decoding system for SGLang, vLLM and an experimental TokenSpeed integration. Its measured target is Qwen3.8-27B with a matching draft model. For teams serving language models, the useful question is whether one drafting system can remain worthwhile as request traffic and hardware change.

What changes in EAGLE-4?

Speculative decoding gives a smaller model the job of proposing upcoming tokens, then asks the main model to verify them. Several accepted tokens can emerge from one verification step. The trade-off is that drafting and checking rejected candidates also take time. A setting that helps a quiet server can become expensive when the same server handles many requests.

EAGLE-4 addresses that trade-off by using measured drafting and verification costs to choose how much speculative work to attempt. The architecture documentation describes a drafter that reads three target-model hidden states, uses sliding-window attention and selects vocabulary candidates through a low-rank head. It then re-scores the shortlisted candidates with the full head rather than scoring every vocabulary entry at full cost.

The serving engines make different scheduling decisions. SGLang can adjust tree depth and verification width using measured costs. In vLLM, verification width follows the batch-size schedule, while the cost model helps decide when to stop increasing draft depth. These are useful distinctions for engineers comparing configurations rather than treating every EAGLE-4 result as the same implementation.

Is EAGLE-4 faster than DFlash 2 and DSpark?

The team's Qwen3.8-27B draft-model card reports 863.8 tokens per second for EAGLE-4, compared with 672.6 for DFlash 2 and 618.6 for DSpark at temperature zero. The conditions are specific: the qualitative split of SPEED-Bench, one concurrent request, four NVIDIA B300 GPUs with tensor parallelism, and a maximum output of 4,096 tokens. All methods use the same inference engine with MegaKernel disabled.

At temperature one, EAGLE-4 reports 608.6 tokens per second versus 560.3 for DFlash 2 and 519.4 for DSpark. That narrower gap matters. The comparison supports a lead in these reported conditions, not a claim that EAGLE-4 wins on every workload or machine. The published checkpoint supplies the drafting component; it is not a standalone chatbot model.

What happens under concurrent traffic?

The measured support matrix provides a more operational picture. On four B300s, SGLang reports a 1.91-times throughput improvement at client concurrency eight and 1.35 times at concurrency 64. On one AMD MI355X, its measured profile reports 2.07 times at concurrency eight and 1.56 times at concurrency 32. Each ratio uses that engine's matched no-speculation control.

These measurements use Qwen3.8-27B in BF16 and temperature-zero SPEED-Bench requests, with two EAGLE-4 server starts. Workload shapes differ between rows, and SGLang's concurrency-64 result queues into 32 running slots. A separate short, high-load vLLM check reaches only 0.970 times an earlier no-speculation control, which was not rerun during that check. Adaptive drafting is therefore not a guarantee of a throughput gain.

Correctness also needs careful wording. The release's checks accept some near-tie or floating-point-tolerance differences, so passing does not mean identical token sequences. Raw launch manifests and request traces remain with the maintainers rather than shipping in the source package. These are team-reported results, not independent reproduction.

What should developers check before adopting it?

Installation requires the project's patched engine builds. Stock SGLang and vLLM packages lack the required hooks, even though the plugin itself does not patch engines at runtime. TokenSpeed remains experimental and carries no claimed throughput result in the release notes. Test the exact serving configuration before treating the headline benchmark as a capacity forecast.

Licensing is another practical constraint. The Eagle 4 Source-Available License is explicitly not an OSI-approved open-source license. Commercial competing offerings require a separate license regardless of revenue, and other commercial use has conditions including a US $20 million aggregate-revenue threshold. Teams should review those terms alongside their hardware and workload fit.

I'm Paras. I help technical founders turn engineering work into clear product stories.

If you're building inference infrastructure, I can help explain the benchmark, the deployment trade-offs and why customers should care without turning a narrow result into a broad promise.

Book a 30-minute call.