DeepSeek-V4.1-Flash Technical Report: Optimized Prefill, Cache, and Decoding
DeepSeek released the technical report for V4.1-Flash, detailing joint optimization of prefill computation, HBM cache, SSD/host persistent cache, and cache invalidation recovery. A notable design is approximate SWA Bounded Replay to replace long-term storage of local cache states.
Key insights
Joint Optimization of Inference Pipeline
DeepSeek-V4.1-Flash's technical report describes a joint optimization across prefill computation, HBM cache, SSD/host persistent cache, and cache invalidation recovery. The most notable design is an approximate SWA Bounded Replay mechanism that replaces the need to store long-term local cache states, reducing memory overhead while maintaining performance.