DSpark is a speculative decoding framework released by DeepSeek (as of June 27, 2026) that significantly accelerates LLM inference by combining parallel token generation with adaptive verification. ### How DSpark Improves Inference Speeds DSpark achieves up to an 80% reduction in latency and 51% to 400% throughput gains (depending on concurrency) by addressing the limitations of previous speculative decoding methods like Eagle3 and DFlash. It uses a **hybrid design**: * **Parallel Block Drafting:** Like DFlash, it generates a block of candidate tokens simultaneously to maximize throughput. * **Sequential Markov Head:** It adds a lightweight sequential head to model token dependencies within the block, which corrects the acceptance rate degradation typically seen in pure parallel drafting. * **Adaptive Confidence Scheduling:** A confidence head evaluates the probability of token acceptance, allowing a hardware-aware scheduler to dynamically adjust the verification length per request. High-confidence prompts receive longer verification blocks, while low-confidence ones receive shorter ones. ### Comparison with Existing Methods * **vs. Eagle3:** Eagle3 uses a learned sequential draft model that provides high accuracy but is limited by its sequential generation, creating a throughput ceiling. DSpark overcomes this by using parallel block drafting. * **vs. DFlash:** DFlash excels at parallel throughput but suffers from degraded acceptance rates at later positions in a block because tokens are generated without knowledge of prior tokens in the same block. DSpark fixes this with its Markov correction head. ### Trade-offs and Considerations * **Training Complexity:** While DSpark-enhanced checkpoints are available for immediate use, training custom draft models using the open-sourced **DeepSpec** stack is resource-intensive. The default training configuration can require up to 38 TB of cache (though this is for training, not inference). * **Inference Overhead:** The system requires additional compute for the draft model and the confidence head, though this is offset by the significant gains in overall generation speed. * **Deployment:** It is designed for production-scale workloads, where its benefits (especially throughput) scale with concurrency. For teams not using DeepSeek V4, the DeepSpec framework allows for training custom draft models for other model families like Qwen3 and Gemma. DeepSeek has open-sourced the **DeepSpec** codebase under an MIT license, providing tools for training and evaluating DSpark, DFlash, and Eagle3. DSpark is currently active in DeepSeek-V4 Flash and Pro production APIs.