On April 30, Xiaomi released the first generative model designed for reasoning - Xiaomo MiMo - on its WeChat official account.
According to publicly available data from Xiaomi, on the public evaluation sets for mathematical reasoning (AIME 24-25) and code competitions (LiveCodeBench v5), MiMo surpassed OpenAI's closed-source reasoning model o1-mini and Alibaba's larger-scale open-source reasoning model QwQ-32B-Preview with only a parameter scale of 7B.
In the same RL training data scenario, MiMo-7B significantly outperformed in the mathematics and code domains, achieving higher scores than DeepSeek-R1-Distill and Qwen2.5-32B.
At the same time, innovations at multiple levels, including data and algorithms during the pre-training and post-training phases, have driven the improvement of MiMo's reasoning capabilities. The core of pre-training is to expose the model to more reasoning patterns, while the core of post-training focuses on efficient and stable reinforcement learning algorithms and frameworks.
Currently, MiMo-7B has been made open source on HuggingFace with four models.
Xiaomi introduced that MiMo represents the initial attempt by the newly established "Xiaomi Large Model Core Team," and stated that "Although 2025 may seem like the latter half of the journey towards large models, we firmly believe that the path to AGI remains long."
Last April, Xiaomi announced the formation of an AI Lab Large Model team, with over 1200 R&D personnel in the AI field. Last December, there were reports that Xiaomi was building its own GPU cluster with ten thousand cards, planning significant investment in AI large models. It was rumored that the large model team had 6500 GPU resources when it was established.
Due to this news, by midday, the stock price of Xiaomi Group rose by 4.74%, with a total market value of 1.29 trillion Hong Kong dollars (approximately 1.21 trillion RMB).
The day before yesterday, Alibaba released the new generation of Qwen3 (Qianwen 3) Tongyi model and made it open source, ranking as the strongest open-source model globally. Tencent announced on the same day that it would restructure its Hunyuan large model research system and increase investment in R&D.
This article is an exclusive piece by Guancha Observer and cannot be reproduced without permission.
Original Source: https://www.toutiao.com/article/7499760085057454632/
Disclaimer: The views expressed in this article are solely those of the author, and your opinions can be expressed through the 'thumbs up/down' buttons below.