DeepSeek's DSpark speeds V4 models 60-85% without architecture changes

DeepSeek has unveiled DSpark, an open-source speculative decoding framework that boosts the generation speed of its DeepSeek-V4 model family by 60–85% while leaving the underlying model architecture untouched. The framework leverages a lightweight draft model to predict multiple tokens in parallel, which are then verified by the main V4 model—drastically reducing sequential computation. According to a paper co-authored by DeepSeek founder Liang Wenfeng, DSpark achieves this acceleration without any accuracy degradation, making it suitable for production deployments where latency is critical. The company has released open-source checkpoints for two variants: DeepSeek-V4-Pro-DSpark and DeepSeek-V4-Flash-DSpark, allowing developers to immediately benefit from the speedup on both high-end and resource-constrained setups.

In addition to the checkpoints, DeepSeek has open-sourced the DeepSpec training toolchain on GitHub, which enables users to train their own speculative decoding draft models tailored to specific tasks or domains. This move lowers the barrier for custom optimization beyond the pre-trained checkpoints. By decoupling inference acceleration from core model retraining, DSpark offers a practical path to faster AI responses without sacrificing model quality or requiring massive additional compute. For enterprises running real-time applications like chatbots, code assistants, or content generation, this translates to lower latency per request and potentially reduced infrastructure costs—all while maintaining the full capabilities of the DeepSeek-V4 family.