DSpark: DeepSeek cuts AI response times with new decoding framework

Chinese AI company DeepSeek has released a new technical framework called DSpark, designed to make its flagship V4 model respond significantly faster. Ben Jiang reports for the South China Morning Post that DSpark increases per-user response speeds by up to 85 per cent, while also reducing the need for powerful and expensive chip infrastructure.

The core problem DSpark addresses is familiar to anyone who has watched an AI tool slowly generate a long response. Conventional AI models produce output one token at a time. When responses are lengthy, this process strains graphics processing units (GPUs) and creates noticeable delays for users. DeepSeek describes this as a “primary bottleneck in serving AI.”

How DSpark works

• A lightweight draft model first proposes candidate responses.
• A larger model then verifies these proposals in batches, rather than one by one.
• A semi-autoregressive method allows the system to produce small chunks of tokens at once, instead of strictly one at a time.
• A confidence-based scheduling system dynamically adjusts how much verification is applied depending on current computing demand.

This combination keeps output quality stable while improving speed. The scheduling system in particular helps balance performance under varying loads.

For content professionals and other non-technical users, the practical implication is straightforward: AI tools built on this technology could feel noticeably more responsive, especially when generating longer texts. For businesses running AI services, the efficiency gains could translate into lower operating costs, as less computing power would be needed to serve the same number of users.

The release reflects a broader shift in the Chinese AI sector. Rather than competing purely on model capability, developers are increasingly focusing on cost-efficiency and user experience as key differentiators.