Xiaomi has launched its first inference open-source large model, Mimo! With 7 billion parameters, it surpasses OpenAI's o1-mini and Alibaba's QwQ-32B-Preview.
Under the same conditions of reinforcement learning (RL) training data, MiMo-7B shows a significantly greater potential in mathematics and coding than other widely used models in the industry, including well-known RL starter models such as DeepSeek-R1-Distill-7B and Qwen2.5-32B.
Chinese humanoid robots are seizing a global $5 trillion "big track"!
Morgan Stanley expects that by 2050, a total of 1 billion humanoid robots will be deployed globally, with annual revenue reaching 4.7 trillion dollars, nearly double the total revenue of the top 20 global Auto Manufacturers in 2024. China's policy support, technological advancements, and manufacturing foundation place it in a leading position in the field of humanoid robots, especially in the Hardware supply chain.
The market may reach a critical juncture in the short term, with Banks and Electrical Utilities showing repeated activity, and the Technology Sector poised to take off.
Track the entire lifecycle of the main Sector.
The pattern of the Chinese smartphone market changed in the first quarter: Xiaomi topped the chart for the first time in ten years, while Apple's market share declined.
① The Analyst believes that the main reasons for Xiaomi's rise to the top are the optimization of its channel system and the national subsidy policy for purchasing devices. ② The overall market fundamentals remain relatively weak, and the shipment volume has not yet recovered to the level of the same period before the pandemic.
Humanoid robot competition: a "technical coming-of-age ceremony" where "flops" and "evolution" occur simultaneously.
① The champion team from Tian Gong, CTO Tang Jian of the Peking Humanoid Robot Innovation Center, told the Star Daily reporter, "The half marathon competition is mainly a limit test for the robot's Hardware and Software motion algorithm, or the stability and reliability of the robot's 'little brain'." ② The Wuxi embodied intelligence sports event opening tomorrow, along with the upcoming robot combat competition held by Yushu, will undoubtedly attract more attention for 'inspection.'
Technology stocks are rebounding, while high-profile stocks are collectively "cooling down". The market may enter a critical period for switching between high and low performances.
Track the entire lifecycle of the main Sector.
Moriarty mcG OP : Xiaomi recently unveiled MiMo, its first open-source large language model (LLM) designed for reasoning tasks, announced on April 30, 2025. Here’s a detailed breakdown based on available information:What is MiMo?MiMo (likely standing for "Mindful Model" or a similar term, though not explicitly defined) is a 7-billion-parameter LLM series, trained from scratch by Xiaomi’s LLM-Core Team for mathematics, coding, and general reasoning tasks. It’s optimized through pre-training and post-training (reinforcement learning, or RL) to enhance reasoning capabilities.The MiMo-7B series includes:MiMo-7B-Base: Pre-trained on ~25 trillion tokens with a multi-token prediction objective to boost performance and inference speed.MiMo-7B-SFT: A supervised fine-tuned version.MiMo-7B-RL: RL-tuned from the base model, excelling in math and code.MiMo-7B-RL-Zero: RL-trained from a cold-start supervised fine-tuned model, achieving 93.6% on specific benchmarks.PerformanceDespite its compact size (7B parameters), MiMo outperforms larger models like OpenAI’s closed-source o1-mini and Alibaba’s Qwen-32B-Preview on key benchmarks (e.g., AIME24, AIME25, LiveCodeBench, MATH500, GPQA-Diamond). For example, MiMo-7B-RL matches o1-mini’s performance in math and coding tasks.It uses a three-stage pre-training data mixture and RL with 130K curated math/code problems, verified by rule-based systems to ensure quality. A test difficulty-driven reward system and data resampling enhance its optimization.AvailabilityMiMo is open-source, with models available on Hugging Face (https://huggingface.co/XiaomiMiMo). Xiaomi supports inference via a forked version of vLLM, though compatibility with other engines is unverified. The team welcomes contributions and feedback at mimo@xiaomi.com.The release includes checkpoints for all model variants, aiming to benefit the broader AI community by providing insights into building reasoning-focused LLMs.SignificanceMiMo marks Xiaomi’s entry into the competitive AI landscape, showcasing its ambition beyond hardware. Posts on X highlight its compact efficiency and superior performance, with sentiments praising Xiaomi’s innovation in open-source AI.Unlike proprietary models, MiMo’s open-source nature allows developers and researchers to adapt and build upon it, potentially accelerating advancements in reasoning-focused AI applications