On January 21st, the Information broke out early in the month, saying:DeepSeek A new generation of flagships will be launched this year, mid-February, during the new calendar year AI Models — DeepSeek V4, which will have greater ability to write code。
On 20 January, on the first anniversary of the release of DeepSeek-R1, developers discovered that DeepSeek had updated a series of FlashMLA codes in GitHub, with 28 files across 114 documents mentioning unknown "MODEL1 " large model identifiers。
This identifier is listed or mentioned separately from the known existing model "V32" (i.e. DeepSeek-V3.2). Based on a code context analysis, MODEL 1 is likely to represent a new model that differs from the existing architecture。
THE DEVELOPERS ANALYSED THE DISTINCTION BETWEEN " MODEL1 " AND " V32 " IN KEY TECHNOLOGIES, MAINLY IN TERMS OF THE LAYOUT OF THE KEY VALUE (KV) CACHE, THE THINNESS OF PROCESSING AND THE DECODED SUPPORT FOR THE FP8 DATA FORMAT. THESE DIFFERENCES INDICATE THAT THE NEW ARCHITECTURE HAS BEEN TAILORED TO THE POTENTIAL FOR MEMORY OPTIMIZATION AND COMPUTING EFFICIENCY。
Previously, the DeepSeek research team also published two technical papers on a new training methodology entitled “Optimizing Disability Connections (mHC)” and a biologically inspired “AI Memory Module (Engram)”. This move leads users to speculate that the new model that DeepSeek is developing has the potential to integrate these latest research results. Please look forward。