Back to News
Tencent•July 29, 2026
AngelSpec
Open Source
Speculative Decoding
LLM Inference
Deep Learning
Tencent Hunyuan
### TL;DR
AngelSpec is a unified, torch-native training framework designed for speculative decoding, supporting both autoregressive Multi-Token Prediction (MTP) and block-parallel drafting architectures. It enables independent scaling of inference and training through a disaggregated architecture, facilitating high-performance model deployment.
Key Insights & Metrics
Pricing
Free
Cost structure
Version
0.1.0
Current release version
Hardware
Linux environment with Python 3.11+ and CUDA 12.4+ on 1 or more NVIDIA GPUs (8 GPUs configured for the default quickstart, and RDMA required only for multi-node setups)
Compute requirements
Category
Open Source
Licensing model
Region
China
Primary region
Key Features
- Unified training pipeline for 6 draft architectures (DFly, DFlash, DFlare, Eagle3, DSpark, MTP)
- Disaggregated training architecture with independent scaling for inference and optimization
- Support for long-context training up to 128k tokens via Ulysses sequence parallelism
- Online evaluation of speculative decoding performance during training
- Document-aware sequence packing with strict cross-document isolation
- Acceptance-aligned training objectives including CE, top-k KL, and D-PACE weighting
Discussion
0
Upvotes
0
Downvotes
0 reviews
Sign in to leave a review
Reviews
No reviews yet. Be the first to review!