Xiaomi Open Sources a 1T Model That Tops Open Weights
2 min readXiaomi is better known for phones and rice cookers than frontier AI, which makes this week’s release harder to ignore. On Tuesday the company open sourced MiMo-V2.6, a trillion-parameter model family that now sits at the top of the open-weights leaderboard, and it put the checkpoints on Hugging Face under an MIT license.
The Open-Weights Race So Far
Open-weight models are those anyone can download, inspect, fine-tune, and run on their own hardware. For the past two years that field has been dominated by Chinese labs: DeepSeek, Alibaba’s Qwen line, Moonshot’s Kimi, and Z AI’s GLM series have traded the top spot between them while Western labs kept their best models closed.
Xiaomi was not part of that conversation. Its MiMo effort was treated as a side project from a consumer electronics company, not a serious challenger.
What Xiaomi Shipped
The MiMo-V2.6 series includes two natively omnimodal models. Pro is the flagship, and Flash is the efficiency tier: a sparse mixture-of-experts design with 309 billion total parameters and 15 billion active at any one time. Per eWeek, a 9B model rounds out the family.
Pro scores 46 on the Artificial Analysis Intelligence Index, first among open-weight models, ahead of Z AI’s GLM-5.3 at 45 and Kimi K3 at 44. On the DeepSWE coding benchmark, Flash jumped from 48.8 to 65.68 and Pro from 58.4 to 72.57.
The training numbers are the part worth staring at. Xiaomi says each model completed 30 reinforcement learning steps in under six days, roughly 750,000 trajectories in total, at a reported cost of $2.62 million for Pro and $850,000 for Flash.
Why It Matters
If those figures hold up, the barrier to reaching the frontier keeps falling. A few million dollars and under a week of compute is within reach of far more organizations than a nine-figure training run, and an MIT license means no usage restrictions at all.
It also widens the gap in strategy. The same week OpenAI cut its API prices in half to defend a paid tier, a Chinese consumer hardware firm gave away a model that tops the open rankings outright. Watch whether Western labs answer with open weights of their own, or keep competing on price.