According to AIbase, as DeepSeek-R1 celebrates its first anniversary, clues about DeepSeek's next-generation flagship model have quietly surfaced. Combined with a report from The Information, this highly anticipated new model (possibly DeepSeek V4) is expected to officially launch as early as mid-February this year (during the Lunar New Year period), and is anticipated to offer even stronger code generation capabilities.

DeepSeek

Developers discovered in DeepSeek's GitHub repository that the updated FlashMLA codebase includes 28 references to a mysterious identifier called "MODEL1" across 114 files. The code logic indicates that "MODEL1" represents a new architecture distinct from the existing "V32" (DeepSeek-V3.2). The key differences between the two lie in the key-value (KV) cache layout, the approach to sparsity handling, and support for FP8 data format decoding, suggesting that the new model has undergone targeted low-level restructuring for improved memory optimization and computational efficiency.

Previously, the DeepSeek team has released technical papers on "optimized residual connections (mHC)" and "AI memory modules inspired by biology (Engram)." The industry generally speculates that these latest research findings are likely to be integrated into the developing "MODEL1," providing core technological support for the upcoming flagship model.