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

When DeepSeek-V4.1 Flash was released, I thought it might just be a post-training iteration version... but after using it for a while, I found it reached nearly 420 Tokens/s in speed, and then Cui said all DeepSeek-V4 Pro models would be taken offline... suddenly I felt this was no small matter... until the Technical Report 《DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression》 [1] was fully released, only then did I realize it should be called DeepSeek-V5 Flash...