Deepseek's DSpark cuts inference lag 85% to defy US chip sanctions
deepseek released dspark, a speculative decoding framework that improves inference speed by up to 85%. this efficiency helps china mitigate the impact of tightening us export controls on high-end ai chips.
deepseek open-sourced dspark, a speculative decoding framework that boosts per-user response speeds by 60-85% by having a lightweight model draft tokens for a larger model to verify in batches. tests on google's gemma and alibaba's qwen models showed dspark beating previous efficiency leaders like eagle3. the technique optimizes gpu utilization and reduces wasted processing on rejected tokens. strategically, this efficiency helps regions like china and the eu achieve higher ai performance with fewer high-end chips, mitigating the impact of us export controls and hardware shortages.