XNews

Tencent Open-Sources AngelSpec Speculative Decoding Framework

Tencent Hunyuan has open-sourced AngelSpec, an end-to-end speculative decoding framework that supports both training and deployment. - AngelSpec includes DFly, a method that delivers 1.98–2.40× end-to-end speedup over autoregressive decoding. - Speedup was measured on the Hy3-A21B model under concurrency levels from 4 to 64. - DFly achieves 10.5–11.8% higher throughput than DFlash. - Training code and Hy3-A21B MTP/DFly drafter weights are publicly available on GitHub, Hugging Face, and ModelScope. - Documentation and a research paper accompany the release.

Key insights

AngelSpec Delivers Significant Inference Speedup for Large Models

Tencent Hunyuan's AngelSpec framework provides an end-to-end speculative decoding pipeline, with their DFly method achieving up to 2.40× end-to-end speedup over autoregressive decoding on the Hy3-A21B model. Across concurrency levels 4 to 64, the speedup remains between 1.98× and 2.40×, and throughput is 10.5–11.8% higher than the DFlash baseline. The release includes training code and drafter weights, making it immediately usable for researchers and engineers aiming to accelerate large model inference.

Source posts