Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

EAGLE-4

Speculative-decoding drafts and a self-tuning harness for SGLang, vLLM and TokenSpeed, under a source-available licence.

Reviewed by ReviewRamble editors · Updated

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

EAGLE-4 is the most carefully measured speculative-decoding release we have read, with 3.42 times faster single requests on SGLang, but it covers one public model, pays off mostly at low concurrency and comes under a licence that bites at US$20 million in revenue.

What is EAGLE-4?

EAGLE-4 speeds up large language model serving with speculative decoding: a small draft model proposes several tokens and the target model checks them in one pass, so the output is the same as the target would produce alone. This repository holds eagle-harness, the package that plugs EAGLE-4 drafts into SGLang, vLLM and TokenSpeed and decides how far to speculate from costs it measures on your machine, so draft trees shrink as the batch grows.

It is published by Eagle AI Corp, which says the work builds on EAGLE, EAGLE-2 and EAGLE-3. The first public draft is for Qwen3.8-27B. The repository was created on 20 October 2026, has about 1,221 stars, ships Docker images and wheels, and released v0.1.5 on 1 November 2026. It is source-available, not open source.

Is EAGLE-4 worth it?

Speculative decoding usually helps a quiet server and hurts a busy one: drafting tokens costs compute that a full batch would rather spend on real work. EAGLE-4's answer is a harness that measures live costs on your hardware and sizes each draft tree to the current batch, so you can leave speculation on. On four NVIDIA B300 GPUs with Qwen3.8-27B, the support matrix reports SGLang at 3.42 times plain decoding for one request, 2.09 at 8 concurrent and 1.39 at 64; vLLM at 2.83, 1.72 and 1.14; TokenSpeed at 3.09, 2.02 and 1.41. On one AMD MI355X, SGLang ran 2.07 times faster at 8 requests and 1.56 at 32.

The documentation is unusually frank about where the gains end. The known-limits page says the README's quick-start SGLang command falls back to plain decoding at high load and measured 1.00 times at 64, 128 and 256 concurrent requests on one B300. A separate high-load recipe on four GPUs keeps pace with plain decoding up to 512 requests. vLLM and TokenSpeed do not switch to plain decoding yet and can be slower than no speculation above 64 requests, so the docs tell you to turn it off there. Correctness is checked against an fp32 rule rather than assumed, and the page lists the prompts where outputs diverge.

Scope is narrow. The measured results cover one public target and draft pair, Qwen3.8-27B, plus Nemotron 3.5 Super, on B300 and MI355X; other architectures, multi-node setups and pipeline parallelism are listed as still being qualified. The quick start downloads about 62 GB and needs a CUDA 13 driver, and pip installs pin a patched SGLang build from Eagle's own wheel index.

The licence is the real decision. The Eagle 4 Source-Available License allows commercial use only below US$20 million in aggregate revenue over any 12 months, forbids competing offerings outright, and counts cloning or forking as acceptance. Eagle AI Corp also controls how accepted contributions are licensed. The speculative-decoding options built into SGLang and vLLM themselves carry none of those strings.

Where it earns its place

  • An inference engineer at a startup running a coding assistant on Qwen3.8-27B

    Cut per-request latency for single users without changing answers

    Serve SGLang with the EAGLE-4 draft and the harness; the support matrix shows 3.42 times plain decoding for one request on four B300s. Free while the company stays under US$20 million revenue.

  • A platform team serving bursty traffic on AMD MI355X

    Speed up quiet hours without slowing peak load

    Use the AMD guide; SGLang measured 2.07 times at 8 requests and 1.56 at 32 on one MI355X. Follow the high-load recipe so busy periods fall back to plain decoding.

  • A researcher comparing speculative-decoding methods

    Benchmark a new drafter against a measured baseline

    Reproduce the support-matrix commands on the SPEED-Bench client at temperature 0, warming the server first, since early requests compile kernels and measure costs.

Before you commit

Setup
Run the published SGLang Docker image, or pip install eagle-harness with the pinned SGLang build in Python 3.12 (needs the CUDA toolkit and a C++ compiler). First start downloads about 62 GB and compiles kernels.
Integrations
SGLang, vLLM and TokenSpeed engines; OpenAI-compatible chat completions from the engine; Hugging Face drafts under EAGLE-Corp; ROCm builds for AMD MI355X.
Data and privacy
Runs entirely on your own GPUs; no hosted service is involved. The Docker example publishes port 30000 on 127.0.0.1 only, and no API key is set in the quick start, so add one before exposing the server.
Cost at scale
No fee below the licence threshold; companies above US$20 million in revenue over 12 months, or offering a competing product, need a commercial agreement whose price is not published.
Limits
Measured on B300 and MI355X only; multi-node and pipeline parallelism not yet qualified; vLLM and TokenSpeed can be slower than plain decoding above 64 concurrent requests; deterministic mode disables speculation at high load.

What’s good

  • Draft trees sized from live measured costs, so speculation can stay on
  • Up to 3.42 times faster single requests with output matching the target
  • Published support matrix with tokens/s, correctness checks and method
  • Works with SGLang, vLLM and TokenSpeed, on NVIDIA and AMD

What’s not

  • Source-available: commercial use capped at US$20 million revenue, no competing offerings
  • Gains shrink to about 1.1 to 1.4 times at 64 concurrent requests
  • One public draft model at launch; other hardware not yet qualified
  • Pins patched engine builds from Eagle's own wheel index

Best for

Teams below the licence's revenue threshold serving Qwen3.8-27B on B300 or MI355X GPUs at low to moderate concurrency.

EAGLE-4 alternatives

EAGLE-4: common questions

Is EAGLE-4 worth it?
EAGLE-4 is the most carefully measured speculative-decoding release we have read, with 3.42 times faster single requests on SGLang, but it covers one public model, pays off mostly at low concurrency and comes under a licence that bites at US$20 million in revenue. We rate it 3.9 out of 5.
How much does EAGLE-4 cost?
Free to use under the Eagle 4 Source-Available License for organisations below US$20 million in revenue over any 12 months and not offering a competing product; above that, a separate commercial licence with unpublished pricing.
What are the downsides of EAGLE-4?
Source-available: commercial use capped at US$20 million revenue, no competing offerings; gains shrink to about 1.1 to 1.4 times at 64 concurrent requests; one public draft model at launch; other hardware not yet qualified; pins patched engine builds from Eagle's own wheel index.
Who is EAGLE-4 best for?
Teams below the licence's revenue threshold serving Qwen3.8-27B on B300 or MI355X GPUs at low to moderate concurrency.
How much does EAGLE-4 cost for a team or in production?
No fee below the licence threshold; companies above US$20 million in revenue over 12 months, or offering a competing product, need a commercial agreement whose price is not published.
Is EAGLE-4 safe for company data?
Runs entirely on your own GPUs; no hosted service is involved. The Docker example publishes port 30000 on 127.0.0.1 only, and no API key is set in the quick start, so add one before exposing the server.
What are the best EAGLE-4 alternatives?
The closest alternatives we have reviewed are SGLang, vLLM and TensorRT-LLM.

Answers checked Nov 21, 2026.