Xiaomi Shrunk 12GB of AI Memory to 2.7GB (MiMo-V3)
MiMo-V3 introduces two techniques—KV bridging and KV reuse—that shrink a million-token KV cache from 12.09 GB to 2.69 GB and cut prefill compute by roughly 5× versus Xiaomi's previous design. Retrieval on a 256K test actually improved, despite the aggressive memory savings. DeepSeek's V4 one flash, released 12 days earlier, independently arrived at the same pair of tricks, both labs building on the same prior work (Yoko and HighSparse). The family is worth watching for agent loops with long tool outputs, but no production weights are out yet and retrieval at a full million tokens remains unverified.