Eagle Launches EAGLE-4, Adaptive Speculative Decoder That Optimizes Inference Across Hardware and Batch Sizes

Eagle released EAGLE-4, an open-source speculative decoding framework designed to accelerate large language model inference while dynamically adapting to batch size, hardware configuration, and workload patterns. The drafter, available for immediate training via TorchSpec and serving through TokenSpeed, represents a refinement of prior EAGLE variants by unifying performance gains across single-request and high-concurrency scenarios without sacrificing inference quality.