Eagle released EAGLE-4, the fourth generation of its speculative decoding method, for teams that need more tokens per second from language models they already serve. A small drafter proposes tokens ahead and the target model verifies them in one pass, keeping the target model's output distribution. EAGLE-Harness then changes draft depth and width as batch size, hardware, and measured cost change, so one trained drafter can cover a single interactive request and a server running hundreds at once. On Alibaba's Qwen3.8-27B in BF16, the launch measurement with TokenSpeed reaches 961.6 tokens per second for one request at temperature 0, on four NVIDIA B300 GPUs with tensor parallelism 4. The paper's matched comparison on the same model and GPUs, with MegaKernel disabled, puts EAGLE-4 at 863.8 tokens per second for one request, 3.59 times vanilla decoding at 240.5 tokens per second. At 512 concurrent requests the same drafter reaches 28,340 aggregate tokens per second, 1.34 times vanilla, 1.70 times DFlash 2, and 1.99 times DSpark. Those fixed-block drafters fall below vanilla decoding at that concurrency. Code and a paper are on GitHub as eagle-harness, with engine plugins for SGLang, vLLM, and TokenSpeed, plus a Qwen3.8-27B draft on Hugging Face. The same writeup reports 2.74 times vanilla throughput on NVIDIA Nemotron 3.5 Super VL for one request and 1.69 times at 512 requests, along with speedups on a single AMD Instinct MI350X, a 4 vCPU cloud instance, and an unmodified phone CPU. Eagle Corp is led by Dr. Hongyang Zhang, a tenure-track professor of AI at the University of Waterloo.