The MiniMax MiMo 3.0 series reconstructs the underlying architecture by sharing and reusing KV cache across layers, reducing prefill computation by approximately 5x and shrinking the KV cache by 4.5x compared to MiMo 2.6 under a 1 million token context.