According to CoinMeta, the team at ByteDance seed discovered that DeepSeek-V4 has a regular weakness in processing long texts. Researchers asked the model to complete a piece of code; although the code itself remained unchanged, just extending the preceding section without any additional instructions caused the model's answers to fluctuate between correct and incorrect. The v4-flash basic version repeats this pattern every 4 token, and a similar phenomenon is also observed in v4.1-flash, with the cycle shortening to every 2 token. ByteDance refers to this periodic variation in retrieval performance as "phase sensitivity." In a retrieval test involving approximately 128,000 token, the accuracy of the v4-flash basic version varied by up to 40.2 percentage points at different positions. After further training, this gap was reduced to 19.1 percentage points, and with additional improvements, it further narrowed to 6.1 percentage points, although periodic fluctuations still exist.