Hangzhou-based DeepSeek, introduced DSpark, a speculative decoding framework for its V4 model family, to speed up inference, or AI response generation, as Chinese developers seek to lower serving costs.

The company said the framework increased per-user response speeds by up to 85% by using a lightweight draft model to suggest candidate responses before a larger model verified them in batches.

The framework also used semi-autoregressive generation to produce small chunks of tokens instead of single tokens, and a confidence-based scheduling system to adjust verification under different computing loads.

The approach is a variant of speculative decoding, an established technique to speed up inference in which a smaller draft model proposes tokens for a larger model to verify.

DeepSeek open-sourced V4 Preview on April 24, 2026, positioning the model family as a cost-effective option for handling 1 million-context inputs.

The company said V4-Pro targets higher performance, while V4-Flash is a faster, cheaper option.

The update comes as Chinese AI companies continue to cut prices for model access and token usage amid intense competition and narrowing capability gaps.

🔗 Source: South China Morning Post