Tencent Open-Sources AngelSpec Speculative Decoding with 2.4x Speedup
Tencent open-sourced AngelSpec, an end-to-end speculative decoding framework. On the Hy3-A21B model, the DFly method achieves 1.98–2.40× end-to-end speedup over autoregressive decoding across concurrency levels 4–64, with 10.5–11.8% higher throughput than DFlash. Training code and MTP/DFly drafter weights are released on GitHub, Hugging Face, and ModelScope.