We noted: Xiaomi promoted 31-year-old Luo Fuli to Level 22, the company's highest rank. She previously worked on V2 at DeepSeek and joined Xiaomi 11 months ago to lead the MiMo team — over the past year, MiMo scaled from 7 billion to trillion-parameter models.
MiMo's cadence this year: open-sourced reasoning model MiMo-7B in April; released V2-Flash (309B total parameters) in December; V2-Pro (total parameters broke the trillion mark) the following March; V2.6-Pro took first place among open-weight models on the Artificial Analysis intelligence index in September, with training livestreamed publicly.
What this is
Three technical points worth noting:
First, a hybrid sliding-window + full-attention architecture (Hybrid SWA — most layers only look at the most recent 128 local tokens, with a few layers seeing global context), which keeps compute costs manageable for 1M-token long context windows (enough to fit a thick book).
Second, Multi-Teacher On-Policy Distillation (multiple specialist models simultaneously teach a single student), letting one model absorb capabilities across domains like code and search.
Third, the reinforcement learning (model learns from environment-provided scores) training dashboard is livestreamed publicly on the official website. The next-generation V3 will use the proprietary HySparse2 architecture.
Beyond the model itself, Xiaomi open-sourced the coding agent MiMo Code (an AI assistant that can autonomously operate software) and the desktop client MiMo Desktop, extending into the product layer.
Industry view
Supporters: open weights, open training, first place on the intelligence index — walking the full path in one year marks the first time a Chinese foundation-model team has openly matched pace with international rivals.
But skepticism remains: Artificial Analysis is a third-party benchmark, weighty but not everything; V2.5-Pro scored only 57.2 on SWE-bench Pro (real GitHub repo fix evaluation), still a gap from top closed-source models; and the next-gen V3's HySparse2 architecture is unproven. These are real uncertainties.
Impact on regular people
For enterprise IT: open weights allow on-prem deployment, but trillion-parameter models don't run on ordinary servers — deployment costs must be calculated upfront.
For working professionals: MiMo Code and MiMo Desktop are pushing model capabilities to the desktop, and workflows like coding and research will be reshuffled.
For the consumer market: Xiaomi's model plugs directly into its full phone-auto-home ecosystem (phones, cars, smart home), making it the closest to ordinary users among this wave of foundation-model companies — but whether the ecosystem actually delivers still comes down to the users.