DeepSeek Open-Sources DSpark and DeepSpec to Speed Model Inference

Overview

• DeepSeek published DSpark, a speculative-decoding framework, plus DeepSpec, a full MIT-licensed toolchain, and released two DSpark-augmented model variants on GitHub and Hugging Face.
• The accompanying paper, credited to Liang Wenfeng with collaborators at Peking University, describes a semi-autoregressive draft generator paired with per-request confidence-scheduled verification to keep more generated tokens acceptable.
• DeepSeek reports that DSpark increased single-user generation speed by 60–85% versus its production baseline (MTP-1) while preserving overall throughput, but those vendor benchmarks have not been independently verified.
• DeepSpec includes data prep, draft-model implementations, training and evaluation scripts so other models such as Qwen3 and Gemma can train compatible draft models and test speculative decoding.
• The release follows a reported major financing round and could push broader industry focus from training to inference engineering by lowering deployment cost and easing real-world adoption of speculative decoding.