Tencent Hunyuan open-sources AngelSpec, a full-stack speculative decoding framework
腾讯混元 · wechat · 2026-07-29
Tencent Hunyuan open-sourced AngelSpec, an end-to-end speculative decoding framework that spans drafter training, architecture design, and online deployment.
- It also released Hy3-A21B MTP and DFlydrafter weights plus training code, making the whole stack reproducible.
- The new DFly drafter reports better acceptance length and throughput than DFlash and MTP, with average speedups of 1.98–2.40× over the AR baseline across 4–64 concurrency, peaking at 2.86× on code/math.
- The D-cut serving policy adds another +15.7% throughput under real high-concurrency traffic by treating verification as a shared batch resource.
- For dialogue, MTP+TTT raises average acceptance rate from 52.8% to 66.4% and acceptance length from 2.58 to 2.99 by fixing train-inference mismatch.
- The framework itself is built on TorchSpec, supports 128k long-context training, and includes real serving-side acceptance evaluation.
Related event: Tencent Open-Sources AngelSpec Speculative Decoding Framework(4 posts)→
More from Infra
- OpenAI cyber eval agents escaped to the internet and compromised Hugging Face infrastructure — nathanbenaich · 2026-10-16
- Redditor Asks: Why No Video Models Built for 24-32GB Consumer VRAM? — BreadUndPeeTears · 2026-10-16
- Rapid-MLX 0.16.2 brings NVIDIA Nemotron 3.5 up to 120B to Apple Silicon Macs — kuchaev · 2026-10-16
- Adobe ships NVIDIA RTX Spark optimizations: up to 2x faster creative apps and 5x Frame.io gains — shashib · 2026-10-16
- Meta's commitments jumped to $349B in a quarter as server memory prices tripled ahead of Oct 21 earnings — tengyanAI · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16