9B 模型“平替”GPT-4o ？！面壁赌对OpenClaw端侧AI，内部上演一人月产65万行代码的效率核爆

Core Insights - The company, Mianbi, shifted its strategy towards edge large models during a competitive landscape in 2023, which faced skepticism until Apple's entry validated their decision [2][22] - Three years later, Mianbi's approach has become more defined, launching the first large model capable of "instant free dialogue" and the AI hardware Pineapple Pi to support full-stack development in hardware scenarios [2][22] Model Development - On February 4, Mianbi released and open-sourced the new flagship multimodal model MiniCPM-o 4.5, which introduces an end-to-end "watch, listen, and speak" capability, allowing for real-time dialogue interactions [3][24] - The model's key innovations include a duplex mechanism where multimodal inputs and outputs do not block each other, enabling continuous perception of audio and video while generating responses [4][25] - The development faced challenges in unified training of various capabilities, requiring a deeper understanding of knowledge absorption and learning dynamics to avoid conflicts between new and existing knowledge [5][25] Performance and Efficiency - The model maintains text capabilities and even achieves slight improvements while ensuring low memory usage and fast response times, providing state-of-the-art multimodal performance with optimal inference efficiency [6][26] - The model's memory is approximately one minute, and it is optimized for low latency, allowing for seamless response generation based on semantic understanding without fixed waiting times [7][28] Ecosystem Development - Mianbi is focusing on building a developer ecosystem to facilitate the deployment of MiniCPM across billions of devices, as relying solely on commercialization is challenging [11][32] - The launch of Pineapple Pi, an AI-native edge intelligent development board, aims to bridge the gap between edge models and applications, facilitating easier development and adaptation [11][34] Competitive Strategy - Mianbi's core philosophy is the "Densing Law," which posits that the knowledge density of large models doubles approximately every 100 days, necessitating continuous innovation to remain competitive [14][35] - The company emphasizes the importance of productization capabilities and infrastructure to extend the competitive advantage of their models in a rapidly evolving market [14][35] Market Positioning - Mianbi believes that the edge market, characterized by diverse applications and terminal types, offers more opportunities for startups compared to the highly competitive general market dominated by large companies [15][36] - The company is focused on addressing core needs in terminal development, aiming for efficiency by achieving strong capabilities with minimal parameters [15][36] Internal Innovation - Mianbi is experiencing a trend towards "one person company" dynamics, where a small team can achieve significant output, reflecting the impact of AI on productivity and collaboration [16][37] - The company seeks to attract AI-native talent who can leverage AI as an intrinsic tool for problem-solving, emphasizing the importance of talent density and quality [17][38] Future Directions - Mianbi envisions a future where edge and cloud collaboration will be the mainstream, with intelligent terminals becoming crucial for real-time data processing and user interaction [18][39] - The company anticipates that as models gain autonomous learning and collaborative capabilities, they will evolve into intelligent agents capable of complex tasks, ultimately leading to a personalized model assistant for every user [20][41]