Title page 1

Contents 4

Foreword 2

Acknowledgements 3

Abstract 6

Résumé 7

Executive summary 8

Introduction 10

1. Delving into AI openness 11

1.1. The term open-source AI is a misleading legacy 11

1.2. Degrees of AI openness: The more model components are publicly released, the easier it is for other actors to reproduce, modify, and use the model 13

1.3. Licensing choices influence access levels, innovation speed, and the potential for beneficial and harmful uses 14

1.4. Clarifying key AI terms: generative AI and foundation models 14

1.5. This report explores the trends, benefits and risks of open-weight foundation models 14

2. Evolution of open-weight models 16

3. Benefits and risks of openly releasing the weights of foundation models 22

3.1. Illustrative benefits 22

3.2. Illustrative risks 23

3.3. Marginal benefits and risks as part of holistic risk assessments 25

4. Conclusions 27

References 28

Tables 4

Table 1.1. Components of the Linux Foundation's Model Openness Framework 12

Figures 5

Figure 2.1. The supply of foundation models has increased consistently, with open-weight models representing over half of commercially available models 16

Figure 2.2. The United States, China and France are at the forefront of open-weight model development, with the largest offerings coming from providers in the US,... 18

Figure 2.3. Over half of foundation model providers are in the United States 19

Figure 2.4. Significant gains in the quality of open-weight models 21

Boxes 5

Box 1.1. Further research is needed to refine and assess open access gradients of AI systems 13

Box 2.1. The AIKoD database on active generative AI foundation models 17