Xiaomi launches MIMO V2 Pro
Xiaomi announced MIMO V2 Pro, a trillion‑parameter model positioned as a strong coding assistant and now testable for free via OpenClaw and KiloClaw. The public testing window gives engineers an easy way to benchmark a new large model variant without enterprise contracts. (x.com)
An anonymous build called “Hunter Alpha” first appeared on OpenRouter on March 11 and, according to Xiaomi, logged more than 1 trillion tokens in calls before the company confirmed the build as an internal test of MiMo‑V2‑Pro on March 18. (mimo.xiaomi.com) Xiaomi’s MiMo‑V2‑Pro documentation lists total parameters at just over 1 trillion, an active‑parameter footprint of 42 billion under its MoE routing, and an ultra‑long context window of 1,000,000 tokens. (mimo.xiaomi.com) The MiMo team says the V2‑Pro increases its hybrid attention MoE routing ratio to 7:1 versus the Flash variant’s 5:1 to balance scale and inference efficiency. (mimo.xiaomi.com) Xiaomi presents benchmark claims showing MiMo‑V2‑Pro outperforming Claude Sonnet 4.6 on coding benchmarks and approaching Claude Opus 4.6 on agent‑style ClawEval tests, and lists an Artificial Analysis Intelligence Index placement of #8 globally and #2 among Chinese LLMs. (mimo.xiaomi.com) Xiaomi announced the MiMo‑V2 family will interoperate with five agent frameworks — including OpenClaw, OpenCode, KiloCode, Cline and BLACKBOX.AI — and the company opened a one‑week global free API access window to accelerate developer testing. (platform.xiaomimimo.com) OpenClaw’s provider docs already include xiaomi/mimo‑v2‑pro as a built‑in model and Kilo’s KiloClaw managed hosting product (GA Feb 24) advertises one‑click OpenClaw deployments and support for hosted agents that can call MiMo models. (docs.openclaw.ai) Xiaomi’s MiMo team is led by Luo Fuli, a former DeepSeek researcher, a detail Reuters cited when the company confirmed Hunter Alpha was an internal MiMo‑V2‑Pro build during the March 18 reveal. (newsbreak.com)
