DeepSeek Launches DSpark to Boost Model Inference Speeds by 85%

DeepSeek has introduced DSpark, a new speculative decoding framework designed to significantly accelerate large language model inference in busy production environments. By utilizing a combination of a parallel draft backbone and a lightweight sequential head, the framework improves per-user generation speeds for DeepSeek-V4 by 60–85% compared to the MTP-1 baseline. The release includes open-source checkpoints and the DeepSpec codebase, which is licensed under MIT for training and evaluating speculative decoding drafters.

Optimizing Inference Throughput

DSpark functions as a serving optimization rather than a new model, attaching a draft module to existing DeepSeek-V4 weights. The system addresses the limitations of previous speculative decoding methods, which often force a trade-off between drafting speed and accuracy. While autoregressive drafters provide strong acceptance but suffer from high costs, and parallel drafters offer low costs but experience rapid acceptance decay, DSpark splits the drafting process into two stages. A parallel backbone generates base logits, while a sequential Markov head adds prefix-dependent bias to maintain high acceptance rates deep into the token block.
This approach allows the system to pull three performance levers simultaneously: drafting faster, drafting more accurately, and verifying smarter. Because the process uses rejection sampling to preserve the target distribution, the output remains lossless, ensuring no degradation in model quality.

Intelligent Verification and Scheduling

To prevent wasted computational capacity during high-traffic periods, DSpark incorporates a confidence-scheduled verification mechanism. A confidence head estimates the likelihood that a draft token will survive verification, while Sequential Temperature Scaling calibrates these scores to reduce expected calibration error from 3–8% down to approximately 1%.
A hardware-aware prefix scheduler further optimizes performance by dynamically adjusting the verification length based on current GPU load. When GPUs are idle, the system verifies more tokens to maximize efficiency; when the system is under heavy load, it reduces the verification budget to protect overall throughput. This load-aware scheduling ensures that the model maintains high performance across various concurrency levels, particularly in structured tasks like code generation and math reasoning where acceptance rates are naturally higher.

Implementation and Availability

The DeepSpec codebase provides a modular environment for data preparation, training, and evaluation. Users can train draft checkpoints against various target models, such as Qwen3 or Gemma4, using provided configuration scripts. The training process freezes the target model, reusing its existing embedding and output head, and employs a total-variation loss to maximize the draft's acceptance rate.
For production deployment, the DSpark-5 configuration—a five-token draft block utilizing the Markov head—is the default implementation. The framework is designed to be integrated into existing serving pipelines without requiring the retraining of the target model, offering a practical path for organizations to increase inference speed while maintaining the accuracy of the original DeepSeek-V4 architecture.