DeepSeek Open-Sources New Framework 'DSpark' That Boosts LLM Inference Speed by Up to 85%

DeepSeek has open-sourced DSpark, a new framework that speeds up LLM responses without degrading model performance. Distributed under the MIT license, this system leverages Speculative Decoding technology to optimize the process by which a model predicts the next word.

DSpark works by having a module that acts as a kind of 'scout,' predicting the path ahead, while the main model quickly verifies it. The more accurate the prediction, the faster the model generates text, and when predictions are inaccurate, unnecessary computation is reduced to maintain efficiency.

This release includes a technical paper along with model checkpoints and DeepSpec, a codebase for training and evaluating speculative decoding systems. This technology was applied to DeepSeek's latest models, DeepSeek-V4-Flash and DeepSeek-V4-Pro, demonstrating the following performance improvements.

• DeepSeek-V4-Flash: 60%~85% improvement in per-user generation speed compared to existing MTP-1
• DeepSeek-V4-Pro: 57%~78% improvement in per-user generation speed compared to existing MTP-1

DSpark is not limited to DeepSeek models and can also be applied to other open-weight model families such as Alibaba's Qwen or Google's Gemma. This means that when companies operate their own models, they can train their own DSpark-style draft modules to reduce inference costs and improve user experience.