DeepSeek AI Releases DeepSpec: Open-Source Framework for Speculative Decoding With Pre-Trained Checkpoints on Hugging Face

DeepSeek AI releases DeepSpec, an open-source speculative decoding framework supporting algorithms like Eagle3 and DSpark, complete with pre-trained checkpoints for models including Qwen3 and Gemma-4-12B now available on Hugging Face under the MIT License.

• DeepSpec is a full-stack open-source codebase released by DeepSeek AI for training and evaluating draft models used in speculative decoding, supporting algorithms including DSpark, DFlash, and Eagle3.
• The framework covers the entire workflow from data preparation and target cache generation to model training across multiple GPUs and evaluation on benchmarks such as GSM8K, MATH500, HumanEval, and others.
• Pre-trained checkpoints for multiple target models including Qwen3-4B, Qwen3-8B, Qwen3-14B, and Gemma-4-12B are publicly available on Hugging Face, with the project released under the MIT License.