DeepSeek open-sources DSpark, a speculative-decoding module for faster DeepSeek-V4 inference

DeepSeek has released DSpark, an add-on speculative-decoding module for its DeepSeek-V4 models, and published it on Hugging Face under an MIT license alongside a companion training codebase called DeepSpec.

What's new

The model card is explicit that this isn't a new base model: "Note: DeepSeek-V4-Pro-DSpark is not a new model. It is the same checkpoint with an additional speculative decoding module attached." In practice, that means existing DeepSeek-V4-Pro and DeepSeek-V4-Flash deployments can bolt on a small drafting head that predicts multiple tokens ahead, which the main model then verifies in a single pass — the standard speculative-decoding pattern used to cut inference latency without changing a model's outputs.

DeepSeek published the release under the MIT license, and paired it with DeepSpec, a full training codebase so outside teams can train their own draft modules against other target models rather than being limited to DeepSeek's own checkpoints. Coverage from VentureBeat and other outlets reported inference speedups in the 60-85% range on DeepSeek-V4-Flash from the technique, though that specific figure sits in DeepSeek's GitHub materials rather than the Hugging Face model card itself.

DSpark builds on the underlying DeepSeek-V4 architecture, which DeepSeek's own technical report describes as using a hybrid attention mechanism where, "In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2" — the architectural efficiency work that DSpark's decoding speedup builds on top of.

Context

Speculative decoding has become one of the standard techniques the industry uses to cut LLM inference costs and latency, typically requiring a separately trained "draft" model paired with the main model. DeepSeek's release continues a pattern the company has followed since DeepSeek-V3: shipping serious inference-efficiency engineering as open, permissively-licensed code rather than keeping it proprietary, which has repeatedly pressured competitors on cost per token. By releasing DeepSpec as a general training codebase rather than a DeepSeek-specific tool, DeepSeek is positioning the technique to be adopted by teams running other open-weight model families entirely.

Why it matters

Inference cost, not just raw model capability, has become a central competitive axis in the open-weight model space, and DSpark is a direct entry in that race: it's a drop-in speedup for models already deployed in production, with no retraining of the base checkpoint required. Releasing the drafting module and its training codebase openly, under MIT, extends DeepSeek's practice of commoditizing infrastructure-level AI efficiency gains — a strategy that has previously forced closed-model providers to respond on price. For any team already running DeepSeek-V4 in production, this is a low-friction way to cut latency and serving costs without touching the model that's already been validated for their use case.