Developers have identified references to an unidentified “MODEL1” in DeepSeek’s GitHub repository, suggesting preparations for a new flagship model. The discovery follows earlier reporting that DeepSeek plans to release its next-generation model, DeepSeek V4, around the Lunar New Year period in mid-February.
Code updates in the FlashMLA library show “MODEL1″ listed alongside “V32,” the identifier for DeepSeek V3.2. Developers noted differences in KV cache layout, sparse processing, and FP8 decoding support, indicating a separate model architecture.
The findings come as DeepSeek’s research team has recently published papers on an optimized residual connection method known as mHC and a bio-inspired memory module called Engram. Some developers have speculated that these techniques may be incorporated into the upcoming model. [TechNode reporting]
Related
