DeepSeek releases DeepSpec for speculative decoding

·2026.06.28 23:18

Key point

DeepSeek released DeepSpec, a tool for training and evaluating models for Speculative Decoding.

Details

DeepSeek-AI released DeepSpec, a full-stack codebase that supports training and evaluation of Draft Models for Speculative Decoding.

DeepSpec includes data preparation utilities, draft model implementations, training code, and evaluation scripts. It currently supports the following three algorithms.

Eagle3
DFlash
DSpark

It also provides various checkpoints for Qwen3 (4B, 8B, 14B) and Gemma-4 (12B) models via HuggingFace, offering an environment for optimizing speculative decoding performance.

open-source deepseek speculative-decoding inference-optimization draft-model deepspec

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.