Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models
Tencent open-sources AngelSpec: a unified framework for speculative decoding that hits 2.4× speedup on 295B models. Here's why inference speed just became a competitive lever.
Why it matters
Speculative decoding is moving from research to production tooling. Tencent's open-source framework and DFly drafter architecture give practitioners a path to 2–2.4× faster inference on large models without retraining — a capability gap closer for teams without frontier-lab resources.
The key facts
6 to knowAngelSpec: open-source torch-native framework for speculative-decoding draft models
Supports six architectures (HY3-295B-A21B confirmed)
DFly drafter: block-diffusion with hybrid target conditioning + hidden-correction autoregressive head
D-cut: runtime-adaptive verification budgeting for inference optimization
1.98–2.40× speedup vs autoregressive decoding on HY3-295B at TP=8, concurrency 4–64
Open-source release enables broader adoption of speculative decoding
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Tencent has released AngelSpec, an open-source torch-native framework for training speculative-decoding draft models across six architectures. It introduces DFly, a block-diffusion drafter with hybrid target conditioning and a hidden-correction autoregressive head, and integrates D-cut for…