DeepSeek Open-Sources DSpark, a Speculative-Decoding Stack for Faster Inference

DeepSeek has released DSpark, a speculative-decoding method that speeds up text generation by roughly 60–85% without touching the underlying model architecture — the gains are purely an inference-serving engineering win. The technique pairs semi-autoregressive draft generation (proposing several tokens at once rather than strictly one at a time) with a confidence-scheduled verification step that decides how aggressively to validate those drafts, cutting the GPU idle time that normally bottlenecks autoregressive decoding. Because it is speculative decoding, the accelerated output is meant to match what standard decoding would produce, so throughput rises without a quality trade-off. DeepSeek reports acceptance-length improvements of about 16–31% over prior draft methods such as Eagle3, with throughput gains ranging from ~50% to several-fold depending on workload.

The more durable contribution is DeepSpec, the open-sourced full-stack codebase that accompanies the paper. It bundles the data-preparation tooling, draft-model implementation, training code, and evaluation harness needed to build and benchmark speculative-decoding models, and it is not locked to DeepSeek’s own checkpoints — it supports families like Qwen and Gemma as well. The shipped DSpark checkpoints are structured as a base model plus an attached draft module, which lets existing deployments bolt on the speedup rather than retrain from scratch.

The significance is less about a single benchmark number and more about lowering the cost of efficient self-hosted inference. By open-sourcing the training and evaluation infrastructure rather than just publishing results, DeepSeek hands the broader community a reproducible path to faster, cheaper serving — continuing the pattern from its earlier open-infra releases of treating inference optimization as shared infrastructure instead of a proprietary moat.