On the first anniversary of DeepSeek-R1's release, the new model "MODEL1" is unveiled

On January 21, according to Quantum Bit, DeepSeek-R1 celebrated its one-year anniversary by revealing a new model “MODEL1”. DeepSeek updated the FlashMLA code on GitHub, with 28 mentions of MODEL1 across 114 files, appearing alongside V32 as a different model. It is known that V32 is DeepSeek-V3.2, and MODEL1 is very likely a new architecture. The specific differences in the code are reflected in KV cache layout, sparsity handling, and FP8 decoding, with multiple variations in memory optimization.