DeepSeek Hints at Upcoming Flagship AI Model Ahead of Lunar New Year
Code references in DeepSeek's GitHub repository point to a new, distinct flagship AI model currently designated as "MODEL1." The discovery aligns with previous reports that the company plans to launch its next-generation model around mid-February.
Developers uncover references to an unidentified "MODEL1" in DeepSeek's GitHub repository, pointing to preparations for a new flagship AI model. This discovery aligns with earlier reports that DeepSeek plans to release its next-generation model, potentially DeepSeek V4, around the Lunar New Year period in mid-February.
Code updates within the FlashMLA library list "MODEL1" alongside "V32," the known identifier for DeepSeek V3.2. Developers observe significant differences in the KV cache layout, sparse processing, and FP8 decoding support, which indicate that this new model features a completely separate architecture rather than a simple iteration.
The findings emerge just as DeepSeek's research team publishes new papers on an optimized residual connection method called mHC and a bio-inspired memory module named Engram. Some developers speculate that these advanced techniques are directly integrated into the upcoming flagship model to boost its performance capabilities.