DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1

DeepSeek open-sourced DSpark, a speculative decoding framework accelerating DeepSeek-V4 per-user generation by 60-85% over MTP-1 losslessly. It utilizes a parallel draft backbone with a lightweight Markov head and confidence-scheduled verification to increase accepted length by 16-31% compared to DFlash and Eagle3. The training repo, DeepSpec, is available under MIT. This optimization enhances AI inference economics, improving efficiency and cost-effectiveness within both centralized and decentralized networks, significantly optimizing overall inference performance for users.