[DigitalToday reporter Chi-gyu Hwang (황치규)] Xiaomi has unveiled its AI model MiMo-V2-Pro with 1 trillion parameters.
According to VentureBeat, MiMo-V2-Pro draws attention because it offers performance close to U.S. AI companies’ models, including those from OpenAI and Anthropic, at about one-sixth to one-seventh of the cost when used for up to 256,000 tokens.
MiMo-V2-Pro is based on a sparse architecture that activates only 42 billion of its 1 trillion parameters during a single inference. It is about three times larger than its predecessor MiMo-V2-Flash and supports a 1 million-token context window. It processes large-scale data with a 7-to-1 hybrid attention ratio. That means it scans 85 percent of the data quickly and applies focused computation to the remaining 15 percent.
According to benchmarking firm Artificial Analysis, MiMo-V2-Pro ranked 10th on the Global Intelligence Index, the highest among China-made models. In GDPval-AA, which measures agentic workplace tasks, it posted an Elo score of 1,426, beating GLM-5 at 1,406 and Kimi K2.5 at 1,283. It fell short of Claude Sonnet 4.6 at 1,633.
Chi-gyu Hwang
delight@d-today.co.kr