The latest release of eagle-hq/eagle-4 is v0.1.3. This repository is the official inference code for EAGLE-4. EAGLE-4 is a speculative decoding tool that adjusts the depth of speculative execution and the tree size according to measured execution cost. The individual items changed in this release are not stated in the provided repository information.
The repository provides libraries and installation and operation guidance for connecting EAGLE-4 to SGLang, vLLM, and TokenSpeed. It tunes the speculation tree to the batch size, hardware, and workload, and also publishes engine plugins, cost correction, and kernel paths. The key point is that it is an implementation connecting research ideas to serving engines. The README reports up to 3.4x generation speed on Qwen3.8-27B while keeping the same answers. Materials introducing the EAGLE-4 release also contained the claim that it is a draft model faster than DFlash2 and DSpark.
However, the up-to-3.4x figure is limited to the measurement environment stated in the README. The speed compared with DFlash2 and DSpark is likewise a claim made in the introducing materials, and no evidence is presented to conclude that the advantage is maintained under other conditions. The license is marked as NOASSERTION, so the usage terms cannot be confirmed from this information alone. Because the individual changes in v0.1.3 are also not given, it is difficult to determine exactly what the release changed.