Leading domestic AI company DeepSeek has been making waves recently. After the one-year anniversary of the release of the R1 model, its technical developments have once again caused a stir among developers. According to the latest GitHub code submission records, in the updated FlashMLA code from DeepSeek, a large number of identifiers pointing to an unknown model "MODEL1" were discovered.
In this codebase spanning hundreds of files, "MODEL1" is mentioned alongside the existing V3.2 version, suggesting that it is not simply an iteration of the current architecture, but likely a completely new model series. Technical details further support this hypothesis: the new architecture shows different design approaches compared to the V3 series in key-value (KV) cache layout, sparse processing logic, and decoding support for FP8 data format, which usually indicates new breakthroughs in computational efficiency and GPU memory optimization.
Previously, it was reported that DeepSeek plans to release a flagship model called DeepSeek V4 during the Lunar New Year of 2026, focusing on stronger coding capabilities. Combined with the recent official publications of two important papers on "optimized residual connections (mHC)" and "AI memory module (Engram)", the outside world generally speculates that "MODEL1" is the engineering implementation of these cutting-edge research results.
