Xiaomi posts new model family
Xiaomi’s Fuli Luo announced a family of models — MiMo‑V2‑Pro, Omni, and a TTS variant — claiming the stack targets the 'Agent era' with 1T parameters and 1M token context; open‑sourcing is promised. That signals more large‑model entrants and new open checkpoints for experimentation. (x.com)
Xiaomi published dedicated product pages for the MiMo‑V2 family on March 18, 2026 and opened developer access through its MiMo platform and documentation portal. (mimo.xiaomi.com (mimo.xiaomi.com)) Xiaomi describes the Pro model as using a mixture‑of‑experts architecture with roughly 42 billion activated parameters during inference to reduce compute compared with a dense configuration. (mimo-v2.org (mimo-v2.org)) The team says the Pro model handles ultra‑long contexts using a hybrid attention scheme reported at a 7:1 ratio to balance memory and throughput for extended multi‑step tasks. (mimo-v2.org (mimo-v2.org)) The flagship was circulated in testing under the codename “Hunter Alpha” and reportedly topped OpenRouter’s daily usage charts before being publicly linked to Xiaomi. (aihola.com (aihola.com)) Xiaomi’s MiMo‑V2‑Omni page publishes benchmark snapshots showing strong audio and multimodal results on MMAU‑Pro and BigBench‑Audio comparisons against recent models from Anthropic and Google in Xiaomi’s charts. (mimo.xiaomi.com (mimo.xiaomi.com)) Xiaomi describes MiMo‑V2‑TTS as trained on a massive speech corpus to support expressive singing and dialects, and the company has made the family available via MiMo Studio and the public API while stating plans to release open‑source artifacts later. (datanorth.ai (datanorth.ai)) (mimo.mi.com (mimo.mi.com))
