China Just Open-Sourced 6 Ways to Speed Up AI Inference (Tencent AngelSpec)
Tencent's AngelSpec is a training framework for speculative decoding that lets you compare six drafting architectures under one config flag, but the repo has been largely ignored (219 stars) while one of its methods, Dflash2, went viral (70k downloads in a week). The paper's core finding is that no single drafter performs best across all workloads: multi-token prediction (MTP) wins on conversational data, block diffusion wins on code and math. AngelSpec's value is not in being the fastest—Dfly, its own method, achieves 1.98–2.40x speedup—but in being the only place you can train all six against your own model and discover which one your traffic needs.