DeepSeek Open-Sources DSpark, Claiming Up to 4x Throughput on V4 via Speculative Decoding

DeepSeek released DSpark, a speculative-decoding method that the lab says raises throughput on its V4 Flash and V4 Pro models by 51 to 400 percent depending on concurrency and use case. The enhanced checkpoints, which pair the original base model with an attached DSpark module, are already published.

Measured against the model's existing multi-token-prediction baseline, DeepSeek reports per-user generation gains of 60 to 85 percent on Flash and 57 to 78 percent on Pro, according to coverage of the release. The shipped configuration, DSpark-5, uses a five-token draft block with a Markov head, and the lab also open-sourced DeepSpec, the codebase for training and evaluating draft models, with reported transfer to other model families such as Gemma and Qwen.

The wide range, 51 to 400 percent, is the part to read carefully. The top of that band reflects favorable batch and concurrency settings rather than a guaranteed result, and the figures are the lab's own. Speculative decoding does not change output quality when verification is exact, so the claim is about cost and latency, not capability. The practical significance is that a frontier-class open-weight model now serves more tokens per accelerator at no cost to quality, which directly lowers the cost of running it outside the major clouds.