DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

DeepSeek-V4.1 Flash introduces major optimizations to reduce KV cache compression, achieving nearly 420 tokens/second by compressing KV cache by 4x while maintaining model quality. The model addresses challenges posed by long-horizon agent workflows through architectural modifications including prefill optimization (using only 8B parameters for prefill vs. 16B for decode) and multi-dimensional KV cache compression techniques such as head count compression, block-based compression, and cross-layer compression with FP4 precision. These innovations aim to overcome storage, computation, and bandwidth bottlenecks that previously limited deployment scalability for extended context processing.