Xiaomi permanently cuts MiMo-V2.5 API prices, reveals full inference optimization

Xiaomi has permanently reduced prices for its MiMo-V2.5 series APIs, with cuts up to 99%. The models use a Hybrid SWA + MoE + multimodal architecture, and a full system overhaul shrinks KVCache storage to about 1/7 of previous levels, slashing inference costs for long sequences.

Key optimizations include separate pools for KVCache management, rebuilt prefix cache trees, a three-tier cache system called GCache, and smarter scheduling. Decode speed gets a boost from expanded memory and speculative decoding, achieving 2.3x speedup for the first 128 tokens. Multimodal inference sees encoder throughput double, and one-hour video processing drops from 156 seconds to 23 seconds.

Xiaomi says the price cut comes from engineering improvements, not reduced model capability. Some optimizations have been shared with the SGLang open-source community, and more open-source plans are on the way.