Tencent Open-Sources an Inference Speedup Stack

Tencent Hunyuan released AngelSpec on July 29 as an open-source framework for training speculative-decoding drafters. First Financial confirmed that the release includes the Hy3-A21B MTP and DFly drafter weights and training code.

What it tries to speed up

Speculative decoding lets a lighter drafter predict future tokens while the target model verifies them in batches. AngelSpec turns that idea into a configurable training and evaluation framework, covering six draft architectures: DFly, DFlash, DFlare, Eagle3, DSpark and MTP.

The technical report says: “We release the AngelSpec framework to support training and extending these speculative-decoding methods.” In practice, the repo is aimed at teams that need reproducible drafter training, online acceptance evaluation and deployment-aware verification.

The numbers are scoped

On Hy3-A21B, the paper reports that DFly raises average accepted length by about 30%, reaches the highest average throughput from concurrency 4 to 64, and delivers a 1.98-2.40x end-to-end speedup over autoregressive decoding. It also reports 10.5-11.8% higher throughput than DFlash.

The GitHub benchmark notes matter: offline throughput uses HY3-295B-A21B with TP=8, temperature 1 and three 120-second windows per cell. Live D-cut traffic is labeled as 8x H20 with concurrency 2-64. That prevents local and online numbers from being blended.

Why it matters

The release moves part of China model competition from parameter counts to serving systems. Agent, coding and long-context workloads are constrained by latency, GPU utilization and tail behavior. If third parties reproduce the gains, AngelSpec could become a practical base for cheaper high-throughput model serving.

Sources: First Financial, Tencent AngelSpec GitHub, arXiv:2607.25852, CocoLoop, Hugging Face; verified release scope, Hy3-A21B TP=8 offline throughput, 8x H20 live-load setup, concurrency 4-64, drafter weights and Apache 2.0 license scope.