DeepSeek releases DSpark: speculative decoding checkpoints for V4-Flash and V4-Pro, open-sources DeepSpec training stack

DeepSeek published DSpark on June 27, 2026 — a speculative decoding framework applied as an add-on module to its flagship DeepSeek-V4-Flash and V4-Pro models, boosting per-user generation speed by 60–85% on V4-Flash and 57–78% on V4-Pro without touching the underlying weights. The enhanced checkpoints — DeepSeek-V4-Flash-DSpark (165B parameters) and DeepSeek-V4-Pro-DSpark (889B parameters) — are now live on Hugging Face.

What's new

DSpark is not a replacement model. As the Hugging Face model card makes clear, it is "the same checkpoint with an additional speculative decoding module attached." The module speeds up inference by speculatively drafting multiple candidate tokens in parallel, then verifying only the promising ones — delivering throughput gains of 51% to over 400% depending on the serving configuration, while keeping output quality unchanged.

Alongside the enhanced checkpoints, DeepSeek open-sourced DeepSpec under the MIT license — the full codebase for training, evaluating, and deploying speculative decoding draft models. DeepSpec ships with three draft model implementations:

• DSpark — a semi-parallel method combining a heavy parallel head with a lightweight sequential Markov head for better token acceptance rates.
• DFlash — DeepSeek's prior speculative decoding approach.
• Eagle3 — based on the widely used Eagle speculative decoding framework.

DeepSpec also supports models beyond DeepSeek's own lineup, including Qwen3 and Gemma target models, and includes data preparation utilities, training code, and evaluation scripts.

Context

Speculative decoding has become one of the most practical techniques for reducing inference latency without degrading model quality. Instead of generating one token at a time, a small "draft" model proposes several tokens; the main model verifies them in a single forward pass. If the draft is mostly correct, the effective throughput can increase dramatically.

DeepSeek has consistently released not just models but the production tooling around them — including quantization techniques and serving infrastructure details. DSpark continues that pattern, shipping the models AND the training stack that produced them.

Why it matters

For teams running DeepSeek V4 in production, a 60–85% reduction in per-user generation latency is material. At the scale DeepSeek's models operate — V4-Pro at 889B parameters — inference cost is often the dominant expense, and latency directly shapes user experience.

The broader significance is the DeepSpec release. By open-sourcing a general-purpose speculative decoding training framework that works with Qwen3 and Gemma, DeepSeek is lowering the barrier for the entire open-source community to apply these techniques to any model. This positions DeepSpec as a potential de facto standard for speculative decoding infrastructure in the open-weight ecosystem.