Xiaomi opens up the inference stack behind MiMo-V2.5

Xiaomi published a technical blog on the inference pipeline behind its MiMo-V2.5 model series, which its own thread points to. The core architectural choice is what Xiaomi calls Hybrid Sliding Window Attention: instead of every layer comparing each new token against the model's entire history, 60 of the model's 70 layers look only at a fixed local window of 128 tokens, and only 10 layers keep a full, global view of the conversation. Because most layers work on that fixed window instead of an ever-growing history, Xiaomi reports this cuts both compute and stored-context (KVCache) size by roughly a factor of seven compared with giving every layer full attention.

Xiaomi frames these as production numbers, not theoretical estimates: it reports server-side KVCache hit rates averaging 93 percent, and 95 percent or higher for its heaviest users, alongside roughly 30 percent higher per-node input throughput, about 40 percent faster prefill processing, and up to 30 percent lower time-to-first-token on long requests, all measured against real production traffic. It also reports a specific multimodal result: encoding a video for the model dropped from 156 seconds to 23 seconds after this optimization work.