AI NewsOpen sourceAnnouncement
DeepTutor v1.7.0 is a self-hostable AI tutor with 48,000 stars
DeepTutor is an Apache-2.0 AI tutor from HKU's Data Science Lab. Version 1.7.0 came out on 28 October. The repo has 48,415 stars.
Image: GitHub
Why it mattersA buyer looking at ChatGPT Study Mode or a paid tutor now has a free, self-hosted option. The tool works with Claude, ChatGPT, Gemini and local models.
A free alternative to a paid AI tutor just put out a big release. HKUDS, the Data Intelligence Lab at the University of Hong Kong, published DeepTutor v1.7.0 on 28 October 2026. The GitHub API reports 48,415 stars for HKUDS/DeepTutor the next day. Trendshift ranked the repository 6th on its daily list the same day. The license is Apache 2.0.
The project describes itself as "lifelong personalized tutoring." The site at deeptutor.info is the authors' hosted version. The GitHub repository is what a user installs on their own machine. The code is Python 3.11 to 3.14 on the backend and Next.js 16 on the frontend. An arXiv paper, 2604.26962, backs the retrieval system.
What v1.7.0 adds
The release notes name five changes. Modular installation lets a user pick which parts to install, instead of taking the whole thing. A refreshed learning workspace reorganises the main interface. Block notebooks with a new editor called Dito let users mix prose with cells. Narrated video lessons generate audio-backed walkthroughs. Resumable embedding checkpoints let a user stop and restart the first index of a knowledge base. No need to start over.
The release page also lists quieter changes under "more reliable tutoring". The authors do not break them down one by one. Earlier October releases, v1.6.14 and v1.6.15, added recoverable indexing, bounded tutoring memory and reliable chat recovery.
What DeepTutor is, in plain terms
A user uploads documents, articles, PDFs, EPUBs or YouTube links. DeepTutor builds a searchable index with retrieval-augmented generation. The user asks questions, writes notes, builds a question bank and runs guided learning paths with mastery gates. The tutor can read the sources page by page and cite them. The release notes list native speech for Gemini. They also add support for several embedding providers. A model picker limits itself to models that match the task.
The project is BYO-model. The release notes mention Claude, ChatGPT via Codex, Gemini, and local models through Lemonade. A long list of regional providers also work, among them Doubao, Zhipu, Bocha, Qianfan, Firecrawl and Aliyun IQS. The user supplies the keys. The code runs locally or in Docker on the user's own box.
Who should look at this
Three groups come to mind. A product manager picking a tutor for a team. A self-learner avoiding a monthly fee. A research lab that cannot send its documents to a cloud. The release cadence is heavy: the authors have cut 32 releases this year, including a v1.7.0 one day after v1.6.15. A buyer who needs the project to sit still for a quarter should pin a version.
A caveat worth stating. The authors publish no accuracy benchmark against ChatGPT or a paid tutor. The star count is a strong popularity signal, not a quality one. Trendshift's daily rank catches the moment, not the long tail. A reader should run DeepTutor on their own material for a week before deciding against a paid alternative.
Source
- HKUDS/DeepTutor v1.7.0, GitHub
- HKUDS/DeepTutor repository, GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.