Researchers find periodic retrieval failures in DeepSeek V4/V4.1 caused by KV-cache compression, with accuracy gaps of up to 40 percentage points

A new paper found that DeepSeek V4 and V4.1 have some issues with long-context retrieval due to KV-cache compression. At 128K tokens, retrieval accuracy varied by up to 40 percentage points depending on where information was positioned. The issue was also present in post-trained models, though the gaps were smaller. Wonder how much this affects actual long conversations and coding tasks... Paper: arxiv.org/abs/2609.36322