DeepSeek's new model MODEL1 exposed

On the occasion of the first anniversary of DeepSeek-R1, the new model "MODEL1" is unveiled. DeepSeek has updated the FlashMLA code on GitHub, with 28 mentions of MODEL1 across 114 files, appearing as a different model from V32. It is known that V32 is DeepSeek-V3.2, and MODEL1 is likely a new architecture. The specific differences in the code are reflected in the KV cache layout, sparsity handling, and FP8 decoding, with several differences in memory optimization. Earlier reports suggest that DeepSeek will release the next generation flagship model around the Chinese New Year in mid-February.