DeepSeek has unveiled DSpark, a speculative decoding framework that significantly speeds up large language model (LLM) inference without requiring any retraining. The system achieves byte-identical outputs while delivering per-user speed improvements of 57–85% on models such as V4, Qwen, and Gemma.
Speculative decoding works by using a small, fast draft model to generate candidate tokens, which are then verified by the larger target model. This approach reduces the number of sequential calls to the large model, cutting latency while preserving output quality. DSpark builds on this concept with optimizations tailored for real-world deployment.
DeepSeek's benchmarks show consistent gains across different model architectures, making DSpark a practical solution for reducing inference costs and improving user experience. The company emphasizes that the method requires no changes to the original model weights or training pipeline.