Written and maintained by CASRAI Editorial Board
Last updated
OpenScholar names two unrelated things. One is a web platform for hosting researcher and institutional websites, which began at Harvard and is now also sold as a hosted product by OpenScholar, LLC. The other is a 2024 research prototype from the Allen Institute for AI (Ai2) and the University of Washington that answers scientific questions with citations drawn from open-access papers. This page separates them so you can tell which one you need. Facts below were checked against each project’s own pages on 2026-11-08.
What it is: the two meanings of OpenScholar
| Dimension | OpenScholar, the web platform | OpenScholar, the Ai2 literature model |
|---|---|---|
| What it does | Builds and hosts research websites and, in the vendor’s current products, “Research Hubs” assembled from existing research sources | Retrieval-augmented language model that synthesizes scientific literature into answers with citations |
| Who is behind it | Origin at Harvard IQSS; the vendor site is copyrighted OpenScholar, LLC; the Drupal project page describes a hosted offering from a software company | University of Washington and Ai2, with collaborators at Meta, Carnegie Mellon, UIUC, Stanford and UNC |
| Audience | Institutions, research offices, individual scholars | Researchers reading and synthesizing literature; AI researchers |
| Status described by its owner | Commercial hosted platform; the Drupal 7 distribution is no longer maintained | “A research prototype” |
Who runs OpenScholar and what it covers
The web platform
The Drupal.org project page calls it “the easiest way to power all of your institution’s research websites.” One installation hosts many self-service sites, with multiple domains, branded themes, drag-and-drop layout and research publishing apps. It began at Harvard, developed by the Institute for Quantitative Social Science (IQSS) with HPAC and HUIT. A 2016 Harvard Gazette article reported that more than 13,000 Harvard members had created more than 6,500 sites, holding over 72,000 publications and more than 300,000 files, and that Princeton, Penn and the Hebrew University were among the other installs. Those are 2016 figures, not current ones.
The vendor site (theopenscholar.com) now describes research infrastructure that “organizes, illuminates, and attracts opportunity for every researcher.” Its main product is Research Hubs, built automatically from existing research sources and then validated by researchers. Other listed products are Match, for funding, and the Web Platform, and institutions can request a 60-day pilot.
The Ai2 literature model
The paper “OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs” (arXiv:2411.14199, first posted 21 November 2024) describes a model that retrieves from a datastore of about 45 million open-access papers and writes citation-backed answers. The Ai2 blog says the datastore draws on Semantic Scholar, with roughly 250 million passage embeddings, and that the underlying peS2o corpus was updated to October 2024. The authors also built ScholarQABench, a benchmark of 2,967 expert-written queries and 208 long-form answers across computer science, physics, neuroscience and biomedicine. They report that their 8B-parameter model outperformed GPT-4o and PaperQA2 on correctness and that its citation accuracy was comparable to human experts. These are the authors’ own results on their own benchmark.
How to find and use each one
Web platform
The vendor’s help center (docs.theopenscholar.com) is organized into these sections: Getting Started, Administrative Settings, Managing Members, Creating Content, Organizing Your Site, Design & Layout, Third-Party Integration, Best Practices and Launch Project. Its popular topics include editing a page, adding a new user, creating menu links, changing the theme, publishing a site and using taxonomy. The sign-in for site administrators is reached from the “Admin Login” link in the footer of theopenscholar.com. A status page is published at status.theopenscholar.com.
Ai2 model
Start from the Ai2 blog post on OpenScholar (allenai.org/blog/openscholar) and the arXiv paper, which link to the released materials. Ai2 states that code, models, the index and data were open-sourced; check the license on each repository before reuse, since the blog does not name one.
Strengths and limits
Web platform
- Strength: a single managed installation lets an institution run many near-identical scholar sites; the Harvard Gazette reported about 99% of scholar sites were structurally identical, which is why one install and one database worked.
- Limit for research offices: the Drupal.org page says the Drupal 7 distribution is no longer maintained and is not covered by the Drupal security advisory policy, so it should be used at your own risk. The Drupal 8, 9 and 10 platform is described as a fully hosted offering from a software company. The page was last updated 6 October 2026.
Ai2 model
- Strength: answers come with citations, and the code, models and data are released openly.
- Limits stated by Ai2: it is a research prototype; it can cite less representative papers, produce unsupported answers or hallucinate citations; and the datastore covers open-access papers only, so paywalled literature is outside its reach. Treat every citation as something to open and read, as you would for any generated reference.
Alternatives and related tools
| Need | Related CASRAI page |
|---|---|
| An open catalogue of scholarly works, authors and institutions to search or reuse | OpenAlex: what it is and how it works |
| A persistent identifier that links a researcher to their works | ORCID service |
| A place to deposit and preserve datasets and outputs | Data repository and trusted digital repository |
| Open publishing infrastructure without author fees | Episciences.org overlay-journal platform and diamond open access |
FAQ
Is OpenScholar the same as the Ai2 model?
No. The Harvard-origin web platform and the Ai2 and University of Washington literature-synthesis model share a name and nothing else in their own documentation.
Is the Harvard software still open source?
The Drupal.org project describes the Drupal 7 distribution as no longer maintained and outside the Drupal security advisory policy, and describes the current Drupal 8, 9 and 10 platform as a fully hosted offering from a software company. The 2016 Harvard Gazette article described the original software as open source and free for Harvard.
Can the Ai2 model replace a literature search?
Ai2 describes it as a research prototype whose datastore holds open-access papers only and which can produce unsupported answers or invalid citations. It can help draft a synthesis, but the sources still need to be read and the search supplemented from databases that cover paywalled work.
Sources
- OpenScholar, LLC, theopenscholar.com and docs.theopenscholar.com, read 2026-11-08.
- Drupal.org, “OpenScholar” project page, drupal.org/project/openscholar, read 2026-11-08.
- Harvard Gazette, “OpenScholar: a Harvard invention,” 12 September 2016, news.harvard.edu/gazette/story/2016/09/openscholar-a-harvard-invention.
- Asai et al., “OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs,” arXiv:2411.14199, read 2026-11-08.
- Allen Institute for AI, “OpenScholar” blog post, allenai.org/blog/openscholar, read 2026-11-08.
