DeepSeek and Peking University open sourced DSpark, an inference framework that raises generation speed by 60% to 85% and can lift single-GPU throughput as much as 6.6 times.
DSpark uses speculative decoding and a semi-autoregressive design so a smaller draft model can propose several tokens before the main model verifies them. The point is to make existing GPUs serve more users rather than wait for new hardware.
What to watch
The story matters because it turns a technical product move into a business and policy signal. The next test is adoption outside the launch circle and whether the numbers hold up in daily use.
Sources: DeepSeek, Peking University, project materials, CocoLoop.