88Z AI, 30 July — Tencent Hunyuan team open-sources speculative decoding framework AngelSpec, covering the full pipeline from drafter training and architecture design to online deployment. It also releases weights and training code for Hy3-A21B's MTP and DFly drafters. DFly achieves SOTA at 4-64 concurrency, accelerating 1.98-2.40x over autoregressive baselines on average, with a peak of 2.86x in code and math. D-cut boosts throughput by an additional 15.7% under high concurrency, while combining MTP with TTT rai
At a glance
- AngelSpec open-sources full-pipeline speculative decoding solution and Hy3-A21B related weights and code
- DFly achieves 1.98-2.40x average acceleration, with MTP combined with TTT raising dialogue acceptance rate to 66.4%

The update
The Tencent Hunyuan team has officially open-sourced AngelSpec, a speculative decoding framework that covers the entire pipeline from drafter model training and underlying architecture design to final online deployment. At the same time, the team released the weights and training code for Hy3-A21B's MTP and DFly drafters for developer use. Performance tests show that DFly achieves state-of-the-art results in the 4 to 64 concurrency range. Compared to traditional autoregressive baselines, it accelerates inference speed by an average of 1.98 to 2.40 times, with a peak acceleration of 2.86 times in code generation and math tasks. Additionally, D-cut technology increases throughput by an extra 15.7% in high-concurrency scenarios, while combining MTP with TTT strategies significantly raises the dialogue acceptance rate from 52.8% to 66.4%.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 腾讯混元Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.