Tencent Hunyuan Releases Open-Source Decoding Framework, Claims 2.4 Times Faster Large Model Inference
Tencent Hunyuan released AngelSpec, an open-source speculative decoding framework for large AI models that supports both training and deployment. Benchmarks on the company’s Hy3-A21B model show the framework delivers a 1.98 to 2.40 times speedup over standard autoregressive decoding across concurrency levels from 4 to 64.
The system also provides 10.5% to 11.8% higher throughput than the DFlash optimization method. Tencent published the training code and model weights on GitHub and Hugging Face, along with a technical paper detailing the architecture and efficiency metrics for developers and researchers.
From the sources (2 posts)
@tencenthunyuan🚀 We’ve open-sourced AngelSpec, an end-to-end speculative decoding framework supporting both training and deployment. On Hy3-A21B, DFly delivers a 1.98–2.40× end-to-end speedup over autoregressive decoding across tested concurrency levels
@teortaxestexRT @TencentHunyuan: 🚀 We’ve open-sourced AngelSpec, an end-to-end speculative decoding framework supporting both training and deployment.…