How OpenAlex assigns topics to works. Every work gets up to three of 4,516 topics, each in a subfield, a field and a domain. Each model version has its own folder with everything needed to replicate it. To learn more about topics in OpenAlex, see the help pages.
- v2 (current, October 2026): a Qwen3-8B classifier trained on 2 million topic choices by Claude Opus 5.5; right on 82% of random works where version 1 was right on 25%. Code, open weights, teacher labels, test sets and every benchmark are in v2/.
- v1 (2024, superseded): multilingual BERT fine-tuned on data labelled by CWTS.
The topics themselves are the same in every version: 4,516 topics with names, keywords, descriptions and a hierarchy, listed here and served by the API. They were built with CWTS (Leiden University); see An open approach for classifying research publications.