According to CoinMeta, Eagle has released EAGLE-4, upgrading its inference and decoding solutions and improving the generation speed of Qwen. The new version focuses on reducing the computational overhead of the draft model and allows for adjustment of the prediction scale based on hardware performance and the number of concurrent tasks. On four NVIDIA B300 GPU, the single-request generation speed of Qwen (3.8-27B) has increased from 240.5 token per second to 863.8 token, a 3.6-fold improvement. When 512 requests are running simultaneously, the total throughput has increased from 21,200 token per second to 28,300, representing a 34% increase. There has also been a significant improvement on mobile devices; on a Motorola Razr with 70 Ultra, Gemma-4-E2B-IT achieved a 3.2-fold increase in long-text generation speed using only CPU, with task completion time reduced from 43 seconds to 14 seconds, and the energy consumption per token decreased by 53%. The code for EAGLE-4 has been made public, supporting integration with SGLang, VLLM, and the experimental TokenSpeed. However, the new version uses a restrictive commercial license, and commercial use requires separate authorization.