EAGLE-4 Team Claims 3.42x Faster Qwen3.8-27B Inference with Speculative Decoding
Summary
EAGLE-4 publishes inference code and eagle-harness, which adds its speculative decoding to SGLang, vLLM and TokenSpeed. The EAGLE-4 Team claims up to 3.42x faster generation on Qwen3.8-27B with four B300 GPUs. The source-available license permits commercial use below $20 million in aggregate revenue.
Key Points
- SGLang with EAGLE-4 drops to 1.39x at 64 concurrent requests on four B300 GPUs, per the repo's support matrix.
- On one AMD MI355X, SGLang with EAGLE-4 runs 2.07x faster at 8 concurrent requests, the EAGLE-4 Team reports.
- The source-available license requires a separate commercial license for competing offerings regardless of revenue.