On September 26, the 2025 Artificial Intelligence Computing Conference was held in Beijing, with the theme of "Building a Foundation, Opening Up, and Burning the Plains".

The conference focused on a number of popular directions in the field of artificial intelligence, among which FlagOS1.5, a unified open source system software technology stack for multiple AI (artificial intelligence) chips, was released, marking that the global AI underlying technology ecosystem is entering a collaborative innovation stage with "open computing" as the core concept.

|FlagOS 1.5 is officially released

The integration of the whole industry chain urgently needs to solve the industry pain points of AI ecosystem fragmentation and difficult software and hardware migration, and open source and opening up are the key to breaking technical barriers and promoting industrial chain collaboration. To this end, Zhiyuan Research Institute and global ecological partners officially released the "FlagOS 1.5 Large Model Full-stack Open Source Technology Base".

With the support of the Beijing Municipal Development and Reform Commission, Zhiyuan Research Institute has taken the lead in cooperating with a number of manufacturers and institutions to jointly create a unified open source system software technology stack FlagOS for multiple AI chips, covering core components such as the efficient parallel training and promotion framework FlagScale, the high-performance operator library FlagGems, and the unified communication library FlagCX, promoting collaborative innovation of underlying technologies in an open source manner, providing an efficient and easy-to-use development environment for "artificial intelligence +", and accelerating technology popularization and application integration.

What is FlagOS?

You can think of FlagOS as a "general-purpose operating system for large models." It is not hardware or an AI model, but a system software toolkit developed by Beijing Zhiyuan Research Institute in conjunction with a number of universities, chip companies and technology companies.

Its goal is clear: to allow large models (such as GPT-like AI) to run smoothly on different brands of AI chips and no longer "pick hardware".

Why is it needed? ——The AI world is too "fragmented".

AI is developing rapidly now, but there is a big problem: each chip company (such as Nvidia, Huawei, Cambrian, etc.) has its own software system, which is not compatible with each other. Like: you have developed an AI model that can run on an A chip; If you change to a B chip, you have to change the code again, or even rewrite, which is time-consuming and labor-intensive. This is called "one chip, one set of tools, one chimney", which seriously hinders the popularization and collaboration of AI.

How does FlagOS solve this problem?

FlagOS is like a "universal translator + universal interface", building a bridge between AI models and various chips. It provides a unified set of software layers; Developers only need to develop a model once and it can run on more than a dozen different AI chips; There is no need for repeated adaptation, which greatly reduces the cost and technical risk of "chip replacement".

It contains three core tools: training inference framework (to make the model run faster), high-performance operator library (to improve computing efficiency), and unified communication library (to make multi-chip collaboration smoother).

Why insist on "open source"? All FlagOS code is open to the public and can be viewed, used, and improved by anyone. There are three major benefits to doing so:

1. Break the monopoly: no longer "stuck" by a certain chip manufacturer;

2. Accelerate innovation: global developers optimize together, and technological progress is faster;

3. Co-build ecology: Chip manufacturers, cloud companies, and research institutions can participate, forming a virtuous circle of "everyone builds together and uses together".

What's new in FlagOS 1.5?

More complete, faster and smarter, so that the newly released FlagOS 1.5 of the AI large model "one run and multiple cores" is like the "general operating system" of the AI world ushering in a major upgrade. It allows large models to no longer be "kidnapped" by a certain chip, can run efficiently on various hardware, and truly achieve "one development, run everywhere".

1. Core capabilities have been comprehensively enhanced

Support more chips for stronger compatibility. Support more than 12 domestic and foreign chip manufacturers; Covering more than 20 global mainstream AI chip models; Whether it is a domestic chip or a foreign chip, it can run with the same system. Truly do: "one set of models, many places".

2. Comprehensive upgrade of key components

3. The utilization rate of computing power has been greatly improved

Higher parallel efficiency through intelligent scheduling and optimization; For the first time, "cross-chip mixed training + heterogeneous reasoning" is realized - chips of different brands can train a model together, greatly improving resource utilization; Provide a more powerful and automated "foundation" for large models with tens of billions and hundreds of billions of parameters.

Investment advice: four main lines, lay out "AI infrastructure dividends" in advance

1. Pay attention to the first batch of domestic chip manufacturers to access FlagOS

It has supported 12+ domestic and foreign chips, and the domestic camp includes: Huawei Ascend, Cambrian, Haiguang, Biren, Moore Thread, etc. It is recommended to pay attention to: Cambrian (688256), Haiguang Information (688041) and other listed companies; An AI chip company that is not listed but has a key ecological niche.

2. Layout of AI system software and tool chain companies

FlagOS promotes the explosion of demand for tools such as "operator generation" and "automatic migration". Pros: AI compiler, high-performance computing library, automated development platform. Attention: AI underlying software companies mentioned by Flush i Wencai; Startups participating in the FlagOS community.

3. Long-term bet on "embodied intelligence" and the robot industry chain

FlagOS has supported the full-link deployment of robot models, indicating that the combination of "AI brain + domestic computing power" will accelerate the landing. It is recommended to pay attention to: robot ontology manufacturers (such as UBTECH, Estun); Agent platform company.

4. Cooperation projects between enterprises and universities participating in the construction of open source ecosystems

FlagOS has 60+ core co-construction units, and participants will gain the right to speak on technology; ecological priority access; Policy and financial support. Attention: Beijing local AI enterprises; The incubation project of the "AI University Public Welfare Tour" partner colleges.