Replication data for: Diffusion as Strategy? Export Controls, Open Weights, and the Global Reach of Chinese AI Models, 2018–2026

Replication package for the Research Article "Diffusion as Strategy? Export Controls, Open Weights, and the Global Reach of Chinese AI Models, 2018–2026" (A. A. Merlet Echavarría, 2026). The package contains a model-level analysis file of 3,034 AI models released between January 2018 and September 2026 (807 Chinese, 1,503 US), built from the Epoch AI "Data on AI Models" database. For each model it records origin, weight accessibility (open or closed), training compute, training hardware and, for derivative models, the country of the base model on which the model was built. The analysis compares the release strategies of Chinese and US developers around the three main rounds of US export controls on advanced AI chips (October 2022, October 2023 and December 2024). Contents:- data/: Epoch AI "All AI Models" and "Notable AI Models" files as downloaded on 24 September 2026 (unmodified); analysis file (models_analysis.csv).- code/: Python scripts that build the analysis file and reproduce all tables and figures (difference-in-differences linear probability models with fixed effects, event study, placebo, logit robustness; pyfixest 0.60.0, statsmodels).- output/: result files (JSON/CSV) and figures (PNG) exactly as reported in the article.- README.md: variable definitions and replication instructions. Main findings: Chinese models became 41 percentage points more likely than US models to be released with open weights by the end of the period, but the divergence emerged only after December 2024 and did not follow the 2022 and 2023 control rounds. The share of derivative models built on a Chinese base model rose from 2.4% before 2023 to 51% in 2025, including 18% of US and 33% of other non-Chinese derivatives. Running the scripts in a clean directory reproduces the analysis file and all tables exactly, and all estimates to within floating-point tolerance. Source and attribution: Epoch AI data are used under CC BY 4.0. Please cite: Epoch AI (2026), "Data on AI Models", https://epoch.ai/data/ai-models. AI use declaration: code drafting, data processing and verification of policy dates and bibliographic records were assisted by Claude (Anthropic). All results are produced by the executable scripts included in this package, and all research decisions were made and reviewed by the author.