DeepSeek and Peking University launch DSpark, boosting AI speed up to 85%

DeepSeek teamed up with Peking University to launch DSpark, a new framework that speeds up large language models in busy production settings. It's already powering preview versions of DeepSeek-V4-Flash and DeepSeek-V4-Pro.

DSpark uses a smart mix of semi-autoregressive design and confidence-based scheduling to generate and check tokens faster. In tests, it outperformed two other popular methods—Eagle3 and DFlash—on models like Qwen3 and Gemma4-12B.

In real-world use on DeepSeek-V4-Flash, DSpark boosted overall throughput by 51% when keeping user speed above 80 tokens per second. On DeepSeek-V4-Pro, it improved throughput by 52% under a 35-token-per-second service level agreement. Matched for throughput, it made individual users 57% to 85% faster.

DeepSeek has open-sourced DSpark's code and models on GitHub, giving developers a practical way to speed up AI without losing quality.