English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Tencent Hunyuan Open-Sources AngelSpec End-to-End Speculative Decoding Framework with 1.98-2.40x Speedup on Hy3-A21B

Forum topic · QianXun · 2026-07-30

Summary

Tencent Hunyuan open-sourced AngelSpec on July 29, an end-to-end (training + deployment) speculative decoding framework, alongside Hy3-A21B MTP/DFly draft model weights. Tested on Hy3-A21B (a 295B/21B MoE flagship model), the DFly approach achieves 1.98-2.40x end-to-end speedup over autoregressive decoding at concurrency levels of 4-64, with throughput 10.5-11.8% higher than DFlash, a competing speculative decoding work. Unlike DeepSeek's DSpark, which focuses on the inference engine, AngelSpec unifies draft model training and server-side deployment in a torch-native pipeline, supporting Multi-Token Prediction (MTP) plus block-parallel training. For AI coding tool vendors such as Cursor and Claude Code, this cuts per-token latency from 30-50ms to 15-25ms, a perceptible improvement for interactive code completion. The release includes training code on GitHub, an arXiv paper, deployment docs on Read the Docs, and draft weights on Hugging Face and ModelScope. AngelSpec is designed to work alongside TorchSpec from LightSeek Foundation for the inference side.

Tencent Hunyuan open-sourced AngelSpec on July 29 — an end-to-end (training + deployment) speculative decoding framework — together with Hy3-A21B MTP/DFly draft model weights. On Hy3-A21B, the DFly scheme achieves a 1.98-2.40x end-to-end speedup over autoregressive decoding (tested at concurrency 4-64), with throughput 10.5-11.8% higher than DFlash, a prominent concurrent speculative decoding work. Unlike DeepSeek DSpark (June 27), which focuses on the inference framework, AngelSpec turns "draft model training + server-side deployment" into a complete engineering pipeline and supports MTP (Multi-Token Prediction) plus block-parallel training. This means AI coding tool vendors can directly reduce response latency — per-token latency drops from 30-50ms to 15-25ms, a perceptible improvement for interactive code completion.

1. Where Speculative Decoding Fits in Engineering

Speculative decoding has become a standard optimization for AI coding tool vendors over the past 12 months. The principle:

1. A small draft model generates k candidate tokens 2. The large target model verifies all k tokens in parallel 3. Correct tokens are accepted, incorrect ones rejected — multiple tokens are produced in a single target-model forward pass

Two engineering pain points have persisted: how to train the draft model and how to integrate it with inference engines. Traditional approaches — Medusa (extra FFN heads), Lookahead (Jacobi iteration), REST (retrieval-based) — all have drawbacks: Medusa requires modifying the main model architecture, Lookahead has unstable speedups, and REST's hit rate depends on caches.

AngelSpec merges both concerns into a torch-native "training + deployment" framework that works out of the box.

2. AngelSpec's Headline Numbers

Measured by Tencent Hunyuan on Hy3-A21B (Hunyuan 3 flagship, 295B/21B MoE):

  • DFly: 1.98-2.40x end-to-end speedup (vs autoregressive decoding) across concurrency 4-64, the common production range
  • DFly throughput 10.5-11.8% higher than DFlash — one of the most watched speculative decoding works of late 2025, reused by Anthropic, Cursor, Codex, and others
  • Supports MTP (Multi-Token Prediction) + block-parallel training — a training objective already used by the GPT-4o family and Llama 4, but rarely unified with server-side deployment in an open-source framework
  • Notably, the AngelSpec paper (arxiv 2607.25852) also acknowledges LightSeek Foundation's TorchSpec, an open-source speculative decoding inference engine. AngelSpec is positioned as the "training side"; for inference, Tencent recommends pairing it with TorchSpec — the two are complementary rather than competing.

    3. Concrete Value for AI Coding Tool Vendors

    For vendors like Cursor, Claude Code, and Codex, AngelSpec's value is not re-proving that speculative decoding works (that consensus is settled), but three concrete things:

    1. Lower per-token latency: Under autoregressive decoding, a 21B-active-parameter MoE model on A100/H100 has ~30-50ms per-token latency. A 1.98-2.40x speedup brings that to 15-25ms — a clearly noticeable difference for interactive code completion. This latency is a core KPI in Cursor/Copilot devtool evaluations. 2. Draft model training is no longer a black box: Previously, companies either trained their own draft models (expensive) or used Medusa-style approaches that require modifying the main model. AngelSpec provides complete training scripts plus directly downloadable Hy3-A21B MTP/DFly draft weights, eliminating self-development costs. 3. Engineering the MTP training objective: Multi-Token Prediction is a core training paradigm for Llama 4 and the GPT-4o family (predicting multiple future tokens at once, accelerating training and improving downstream performance), but few open-source inference engines support MTP training. AngelSpec fills this gap — Chinese AI coding tools can now more aggressively combine MTP training with speculative decoding.

    4. Context: Competing Work Since July

    Recent output on the AI inference infrastructure side:

  • 6/27: DeepSeek DSpark speculative decoding framework open-sourced (decoding engine focus)
  • 7/03: Alibaba DAMO Academy Elements Claw 1B superconductor-material discovery model + Agent framework
  • 7/12: ByteDance Claude Code desktop built-in browser
  • 7/23: Cursor Router cache-aware model routing + $4.63 per commit vs Fable 5's $12.69
  • 7/29: Tencent Hunyuan AngelSpec (unified training + deployment framework)
  • Chinese tech giants' consecutive output in AI inference infrastructure now forms a de facto standards cluster: DeepSeek on inference engines, Alibaba on training paradigms + tools, ByteDance on IDE integration, Tencent on speculative decoding training frameworks. Each piece alone is not "disruptive," but together they give the Chinese AI coding ecosystem a path independent of the NVIDIA CUDA moat and of OpenAI/Anthropic APIs.

    5. The Open-Source Release

    Tencent Hunyuan's release is comprehensive, with five entry points:

  • GitHub: https://github.com/Tencent/AngelSpec (training code)
  • arXiv paper: https://arxiv.org/abs/2607.25852 (DFly algorithm details)
  • Read the Docs: https://angelspec.readthedocs.io (deployment docs)
  • Hugging Face: https://huggingface.co/collections/AngelSlim/angelspec (Hy3-A21B MTP/DFly draft model weights)
  • ModelScope: https://modelscope.cn/collections/AngelSlim/AngelSpec (China mirror)
  • Delivering weights, code, paper, docs, and a domestic mirror simultaneously is the "standard move" for Chinese major-lab open-sourcing in 2026 H2 — minimizing the barrier to adoption, unlike some 2023-era Chinese releases that published papers without weights.

    6. Potential Impact on Cursor / Anthropic / OpenAI

    Speculative decoding is fundamentally about extracting more tokens per target-model forward pass; whoever has the more accurate draft model and tighter integration gets lower inference costs. Anthropic's Claude Fable 5 system reportedly uses speculative decoding internally (specifics undisclosed), and OpenAI's GPT-5.6 Sol disclosure of "15%+ further token generation efficiency gains" likely also involves a proprietary speculative decoding scheme.

    If AngelSpec sees broad adoption among downstream tool vendors in 2026 H2:

  • Chinese AI coding tools: per-commit inference costs can drop another 30-50% (stacked with Cursor Router's cache-aware routing)
  • International AI coding tools: either adopt similar approaches or lose competitiveness on price/latency
  • Chips: speculative decoding is slightly less dependent on H100/H200 than autoregressive decoding (the draft model can run on smaller GPUs); adapting AngelSpec to domestic chips (Huawei Ascend, Cambricon, Hygon) may be easier than running the main model directly — a hidden benefit for domestic compute
  • 7. Original Links

  • Tencent Hunyuan X announcement: https://x.com/TencentHunyuan/status/2082447023626944936
  • GitHub repository: https://github.com/Tencent/AngelSpec
  • arXiv paper: https://arxiv.org/abs/2607.25852
  • Hugging Face collection: https://huggingface.co/collections/AngelSlim/angelspec
  • Read the Docs: https://angelspec.readthedocs.io

Tags

#tencent-hunyuan#angelspec#speculative-decoding#llm-inference#open-source#mtp#ai-coding-tools#hy3-a21b

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178503807