XiaomiMiMo releases MiMo-V2.5-Pro, an open-source 1.02 trillion parameter MoE language model
The AMW Read
XiaomiMiMo enters the trillion-parameter open-weight arena, meaningfully updating China's foundation-model landscape and scaling-law dynamics.
XiaomiMiMo releases MiMo-V2.5-Pro, an open-source 1.02 trillion parameter MoE language model
China's XiaomiMiMo has released MiMo-V2.5-Pro, an open-source mixture-of-experts model with 1.02 trillion total parameters and a 1 million-token context window. The model is available on Hugging Face. According to the model card, MiMo-V2.5-Pro was developed by the Xiaomi MiMo Team and is designed to comply with Chinese regulations.
Why it matters: This release validates the open-weight mega-model trend, where Chinese players openly distribute frontier-scale models—a pattern that reshapes the economics of foundation model availability. With 1T parameters and a 1M context window, XiaomiMiMo enters the same weight class as DeepSeek and Qwen, intensifying the open-source arms race and exerting downward pricing pressure on proprietary API pricing globally.
Expert take: XiaomiMiMo's entry into the trillion-parameter club signals that the compute and capital barriers to building frontier models are falling faster than expected, especially for Chinese labs with access to domestic GPU supply chains. The open-weight distribution strategy mirrors DeepSeek's move to build ecosystem mindshare, which may accelerate commoditization of foundation model tiers. However, regulatory compliance with Chinese law could limit enterprise adoption outside China.