Executive Summary
Recently, the open source ecosystem has been drawing significant attention as it provides
the technological foundation for innovation in artificial intelligence (AI) and for AI
transformation (AX, AI Transformation). Among 28 major deep learning frameworks, 25 are
open source technologies, which have improved the ease of AI development and accelerated
the scaling-up of models. In addition, open source models such as the Transformer (the
architecture underlying LLMs) and BERT (which established the pre-training paradigm) have
enabled the rapid diffusion of AI technology, giving rise to the new concept of Open Source
AI.
However, in order to address the confusion caused by the absence of a clear concept of
Open Source AI, the Open Source Initiative (OSI) published the Open Source AI Definition,
and the Linux Foundation released the Model Openness Framework (MOF). Both frameworks
commonly define Open Source AI as AI that permits four freedoms (use, study, modify, and
share) and distinguish the scope of disclosure into data, the model, and software (code).
Although there are some differences between the two, they are meaningful in that they
present clearer concepts of Open Source AI and open models.
Looking at recent Open Source AI trends, AI-related projects on GitHub and open model
development centered on Hugging Face are increasing rapidly. As of 2024, the number of AI
projects on GitHub reached 4.32 million, and as of December 2025, the number of open
source models on Hugging Face surpassed 2.25 million. In addition, 89% of companies use
open source technologies in the AI development process, and 63% use open models. The
main reasons are innovation (67%), market standards (67%), productivity (50%), and
development cost reduction (49%).
According to Epoch AI’s analysis of Notable AI Models, releases of open source models
have increased since 2018, and 47.3% of notable AI models released after 2018 were open
source. The countries leading open model releases are the United States and China, and the
majority of participating organizations are from industry and academia. The primary types of
released models are language models, vision models, and multimodal models. Behind the
expansion of open model releases lies the U.S.–China competition for AI leadership. While
the early Open Source AI ecosystem was led largely by U.S. companies (such as Meta and
Google), since 2025 Chinese companies such as DeepSeek and Alibaba have been rapidly