Tencent's Hunyuan team has open-sourced AngelSpec, a speculative decoding framework.

Tencent's Hunyuan team has open-sourced the speculative decoding framework AngelSpec, covering the entire chain from drafter training and architecture design to online deployment. Simultaneously, they have open-sourced the MTP and DFly drafter weights and training code for Hy3-A21B. DFly achieved state-of-the-art performance (SOTA) at 4 to 64 concurrency levels, with an average speedup of 1.98–2.40 times compared to the autoregressive baseline, and a peak mathematical speedup of 2.86 times. D-cut further improved throughput by 15.7% in high concurrency, and the combination of MTP and TTT increased the conversation acceptance rate from 52.8% to 66.4%.