DeepSeek ships DSpark and open-sources its DeepSpec speculative-decoding stack

DeepSeek shipped DSpark to production and open-sourced DeepSpec, its full training stack for speculative decoding, on June 27, 2026. The company says the technique speeds per-user text generation by 60 to 85 percent on its DeepSeek-V4-Flash model against the prior multi-token prediction baseline.

Speculative decoding is an inference technique in which a small, fast draft model proposes tokens that the larger model then verifies in parallel, cutting latency without changing outputs. DeepSeek’s DSpark, which it describes as confidence-scheduled speculative decoding with semi-autoregressive generation, is now live on DeepSeek-V4-Flash and V4-Pro, the company’s 284-billion- and 1.6-trillion-parameter models.

The more consequential move for other labs is the open release. DeepSpec, published on GitHub under an MIT license, bundles data-preparation scripts, multi-GPU training pipelines and evaluation across nine benchmarks, plus three algorithms: DSpark, DFlash and Eagle3. A default run needs a single eight-GPU node and roughly 38 terabytes of storage.

DeepSeek reports that tuning confidence thresholds lifted its chat-prompt acceptance rate from 45.7 percent to 95.7 percent, and math-reasoning acceptance from 76.9 percent to 92.5 percent. In offline tests, the company says DSpark beats Eagle3 by 26.7 to 30.9 percent on accepted length across the Qwen3 family at 4-billion- to 14-billion-parameter scale.

Every one of those numbers is vendor-stated and has not been independently benchmarked. Acceptance rates and throughput gains depend heavily on workload, and DeepSeek’s reported throughput improvements span a wide 51 to 400 percent depending on load, a range that makes a single headline figure misleading.

Because the training code is open, outside groups can now reproduce or refute the claims on their own hardware, which will determine whether DSpark becomes a standard or stays a DeepSeek-specific result.