On June 27, DeepSeek and Peking University released DSpark (Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation), with founder Liang Wenfeng listed as a co-author. This is not another "main-model weight upgrade" — it's an engineering-grade breakthrough targeting the online LLM inference system bottleneck: on the production V4-Flash / V4-Pro engine, DSpark improves per-user generation speed by 60%–85% at the same throughput versus the in-service MTP-1 baseline, and at strict 120 TPS interactive SLAs it raises usable system throughput by 6.6×.

DSpark's solution has two layers. On the drafter side, the body uses a parallel backbone, with a lightweight sequential block (Markov head or RNN head) at the tail to recover local dependencies inside the block — preserving the throughput of the parallel drafter while keeping suffix decay under control. On the verifier side, a confidence head estimates the per-position prefix-survival probability, combined with a hardware-queue-aware scheduler that dynamically decides how long to verify this round, directing target compute at the tokens most likely to be accepted.

The experiments align training data and drafter depth on Qwen3-{4B, 8B, 14B} and Gemma4-12B, and compare DSpark head-to-head with Eagle3 (autoregressive) and DFlash (parallel): on Qwen3-4B/8B/14B, the macro-average accepted length improves 30.9% / 26.7% / 30.0% over Eagle3 and 16.3% / 18.4% / 18.3% over DFlash. All three benchmark families (math, code, daily chat) move up, not just one task.

More important is production deployment: at a moderate 80 tok/s/user SLA, DSpark lifts V4-Flash aggregate throughput by 51%; at a strict 120 tok/s/user SLA where the MTP-1 baseline is already at the concurrency limit, DSpark still pulls 6.6× usable throughput. In other words, DSpark pushes the "latency-throughput" Pareto frontier outward as a whole, making the "compress concurrency to preserve interactive latency" high-sensitivity scenario a default option for the first time.

The companion open-source DeepSpec is another highlight: it unifies the training and evaluation code for DSpark, DFlash, and Eagle3 in one full-stack repo under MIT, with 12 trained drafter checkpoints for the Qwen3 and Gemma4 series released on Hugging Face. Downstream teams can take them for domain adaptation directly — and that's the key step for DSpark to become "industry infrastructure."