This page may contain stale information. Last updated: 2026-08-01
Definition
Inference acceleration technique where a cheaper draft model proposes tokens that a target model verifies in parallel, improving throughput without changing final distribution (under correct algorithms).
Key Points
- DeepSeek’s dspark is cited as part of deepseek-v4-flash 0731 serving stack
- Complements post-training gains for agentic/coding latency