Xiaomi explains how it serves MiMo-V2.5 models
Xiaomi has published an engineering write-up on serving its MiMo-V2.5 models, showing how hybrid sliding-window attention cuts KVCache to about one-seventh of full attention and what it took to deliver those savings in production.
Xiaomi's MiMo team published a detailed account of its inference stack for the MiMo-V2.5 series, which includes MiMo-V2.5 and MiMo-V2.5-Pro. The page is titled as a launch of three MiMo-V2.6 models, but the text available describes no V2.6 models, and no names, specs or availability details are given.
The write-up centres on Hybrid Sliding Window Attention (SWA). MiMo-V2.5-Pro has 70 layers: 10 use full attention and 60 use SWA with a window of 128 tokens. Because six-sevenths of the layers attend only within a local window, Xiaomi says attention compute and KVCache storage both fall to roughly one-seventh of a full-attention design. That lowers prefill cost, which is mostly limited by compute, and decode cost on long sequences, which is mostly limited by memory bandwidth.
By Xiaomi's own estimate, MiMo-V2.5 and V2.5-Pro have the second-smallest KVCache among Chinese models, behind only DeepSeek-V4-Pro and Flash. The team notes that fixed overheads mean real costs don't fall exactly in line with cache size. The savings are similar on short inputs and grow with context length.
The team says these theoretical gains don't carry over to production automatically. Hybrid SWA makes cache hits, prefix matching and keeping full-attention and SWA layers consistent much harder. At launch, mainstream open-source inference frameworks and caching systems didn't fully support SWA, so Xiaomi started from SGLang v0.5.5 and rebuilt its KVCache system. The post also covers tiered caching, an SWA-aware prefix cache tree, scheduling, the prefill and decode pipelines, and multimodal encoder throughput.
The material available cuts off partway through, so the full set of results isn't included here.