Over a year ago, ChatGPT-3.5 made its debut, with its breakthrough natural language interaction capabilities shaking the world. In the following dozen months, the Transformer architecture and large language models (LLM) became epoch-making technologies, sparking entrepreneurship and innovation across the entire AGI field. OpenAI and Google successively released "AI toolkits" such as GPT-4o and Project Astra, ByteDance launched the "Doubao" large model family, Tencent introduced the "HunYuan" large model, and the AI application field flourished, entering a new stage of major development. Additionally, technologies such as digital humans, digital twins, life sciences, embodied intelligence, and the metaverse are gradually entering people's field of vision. The era of AI-Native is calling us, and for enterprises, researching and applying AI technologies has become a "must-have" for digital and even intelligent transformation.
However, on the prosperous side, we also clearly see that the Transformer architecture highly relies on the "scaling law". In addition to the traditional elements of computation power, data, and algorithms relied upon by deep learning, electricity has also joined in, becoming a new bottleneck for enterprises, data scientists, and application developers. High electricity consumption and energy costs pose challenges to the infrastructure built for the era of cloud computing; expensive computational costs and cumbersome massive data processing slow down the pace of application innovation; best practices for AI applications in scenario exploration, development, deployment, and operation are still being explored...
The official release of MatrixOS
To address the challenges of the AI Big Model era, Matrix Origin has launched the MatrixOS product. MatrixOS is an open-source AI-Native operating system that connects computing power, data, knowledge, models, and enterprise applications, providing a comprehensive end-to-end AI Stack service framework.
MatrixOS adheres to the open concept, embracing open-source technologies. It is built on a containerized architecture that is highly modular and scalable. It features a powerful heterogeneous data storage and processing platform, along with various open-source large-scale models, finely-tuned applications, and orchestration frameworks. MatrixOS consists of three core sub-products, which can either be combined into an all-in-one solution or independently deployed to provide services.
MatrixDC: Computing Power Service Platform
MatrixDC is a software product for heterogeneous computing power management and scheduling. Serving as the computational foundation of MatrixOS, it features modular, scalable, and high-performance cloud-native service capabilities. It offers enterprises a range of platform capabilities including heterogeneous computing pool scheduling, ultra-large-scale computing clusters, and intelligent operational service quality assurance. Through flexible billing models and higher cost-effectiveness, it meets diverse customer needs. It provides a ready-to-use distributed computing power pool, offering a fast, stable, efficient, and flexible distributed support environment for data processing, training, fine-tuning, and inference. Targeting developers, it provides comprehensive development API/SDK, helping enterprises quickly integrate with the MatrixDC platform and achieve their desired business goals.
Additionally, MatrixDC supports deep integration with NVIDIA AI Enterprise and OminiVerse software platform, coupled with comprehensive expert technical support, to provide customers with AI application development, model training, inference, and other full-lifecycle management services, empowering enterprises with AI capabilities. Moreover, MatrixDC will gradually support the integration, networking, and computing services of domestic GPU chips.
MatrixOne: Hyper-converged Data Management Platform
MatrixOne is a hyper-converged data management platform. As the data processing layer of MatrixOS, it is designed for cloud-native and containerized environments, employing a storage-compute separation architecture. It supports various heterogeneous workloads such as OLTP, OLAP, time-series, stream processing, machine learning, and handles multiple data types. Its storage layer, based on shared object storage, enables low-cost storage of massive data and facilitates collaboration. The computing layer, based on stateless containerization, allows rapid elastic scaling to cope with fluctuations in workloads. Developers can quickly and comprehensively build business systems and data analysis applications with MatrixOne. For LLM (Large Language Model) scenarios, its vector capabilities enable the rapid construction of knowledge bases based on contextual data understanding. MatrixOne is also a fully open-source project, and we warmly welcome community developers to join and contribute.
MatrixGenesis: AI Intelligent Agent Development Platform
MatrixGenesis, as the core application development layer of MatrixOS, is leading enterprise-level AI applications into a new era. This innovative platform not only covers the entire development process of large models but also provides full lifecycle development support from model selection, deployment, inference services, fine-tuning, to real-time integration with structured system data. MatrixGenesis is committed to providing end-to-end development process experience for enterprises, ensuring seamless integration at every step to accelerate the development and deployment of AI applications.
The MatrixGenesis development platform is designed specifically for AI developers, aiming to provide an efficient, flexible toolchain and platform to support rapid iteration of large model applications from scratch. Whether based on existing base models or user-selected models, users can develop intelligent applications that fit their application scenarios. MatrixGenesis covers the full range of development needs from Model Finetune (DPO, PPO), Model Alignment, and Model Evaluation, to the establishment of knowledge bases, knowledge graphs, and low-code construction of Multi-Agent Workflows. Additionally, we offer RAG/Prompt tuning and evaluation for specific scenarios, helping developers iterate quickly and continuously improve the actual AI capabilities of their applications.
At the same time, MatrixGenesis will also collaborate with partners to create open models, application stores, and intelligence-sharing subscription capabilities, allowing more developers to participate in the ecosystem construction of intelligent applications, quickly gather users for their intelligent applications, and better serve the users they face.
Looking forward to the future: openness, cooperation and innovation.
Based on the capabilities of MatrixOS and 21vianet AIDC Wanka cluster, the AI-native cloud platform neolink.AI is also set to be released soon, marking the first large-scale implementation of MatrixOS in the industry.
The release of MatrixOS is not only a bold prediction and proactive layout of MatrixGenesis for the future of AI but also a sincere invitation to the entire technology community. We welcome more AIDC providers, large model manufacturers, and data intelligence application vendors to join this open and open-source ecosystem. We firmly believe that through open collaboration and shared innovation, we can collectively promote the advancement of AI technology and create greater value for enterprises and society.