A Peking University education model that got 650 stars in 52 days on the back of "#1 on HF's daily chart" and a Datawhale press piece

haolpku/Omni-Edu is an open-source K-12 education model from Peking University's DCAI team, in 4B, 9B and 27B sizes, with weights, training data and evaluation code all released. It went from zero to 653 stars on GitHub in 52 days, gained +247 in the last week, and is still accelerating. Very few people actually download it to use it: the official weights on Hugging Face total about 970 downloads over 30 days, and the PyPI package gets 61 downloads a week. Its stars come from academic prestige and Chinese media reposts, not from users. Below: what it is, where the stars came from, and what an indie developer outside a top lab can take from it.

One WeChat article, and the next day it had gained 52 stars

On the noon of Oct 2026 29 13:34, the WeChat Official Account of the AI open-source community Datawhale published an article credited to "Peking University OpenDCAI Team," titled "#1 on the Daily Chart, Peking University's Major Open-Source SOTA-Level Education Model OmniEdu!" It opened with an old saying, "know the what but not the why," then asked: "When AI first has nearly unlimited knowledge, we should ask even more: what does it actually mean to teach a person?" The third paragraph released the number that spread best: compared with five closed-source models including GPT-5.4, Claude-Opus-5 and Kimi-K3, OmniEdu-27B "won 51 out of 60 comparisons."

The same day, the article appeared on Sohu (two accounts), a Zhihu column and the TRAE Chinese community, a typical multi-platform WeChat syndication. The repo also merged a PR that day adding a community group QR code to the README. The week before, it had been gaining a dozen or twenty stars a day; on Oct 30 it gained 52 in one day (daily star counts from GitHub).

What it is

A set of open-source models that "can teach." The base is Qwen3.5 / Qwen3.8, and the authors did full-parameter fine-tuning on about 70k teaching samples. The README splits the target abilities into four parts: solving problems, aligning with curriculum standards, identifying where a student is stuck, and knowing how to guide the next step. The paper's framing is "from Solver to Tutor," meaning from only solving problems to acting as a teacher.

The release is complete: each of the three sizes comes in BF16 and FP8 versions; there is also training data, a data processing pipeline, training entry points, an evaluation suite covering 12 benchmarks, and a pip package called omniedu. Community-made GGUF quantized versions total about 20k downloads (Hugging Face, 11-06), far more than the official weights. Those who want to run it locally seem to mostly download the quantized versions.

The lead author, Liang Hao, is a Peking University PhD student. His earlier work in the same direction, K12-KGraph, was accepted at NeurIPS 2026 and has 458 stars. His lab, OpenDCAI, already has older repos such as DataFlow (8.4k stars) and DataFlex (3.4k), so it had a ready-made following.

How the stars came in waves

(GitHub daily star counts; 11-06 totals 653 stars and 103 forks.)

Each acceleration lines up with an exposure event: first the paper charts, then the media article, then the directories. The English-speaking side has only two HuggingPapers posts and one Reddit thread, no Hacker News, and no big accounts sharing it. The fork ratio is 16%, which is on the high side and suggests students and researchers forking it to run evaluations. The curve climbs day by day with no single-day spike, so no sign of star-farming (inference; I couldn't check the starring accounts).

Why this one

1. Tag-friendly. "Peking University," "#1 on the daily chart," "SOTA," "beats GPT-5.4 and Claude": it hits every term Chinese AI media loves, and the headline is basically those words strung together.

2. Rode the policy. In 2025 April, the Ministry of Education and eight other departments issued a notice calling for "building self-controlled, education-specific large models." Universities, education tech companies and school IT teams are all watching this, and an open, locally deployable model trained on the Chinese curriculum standards fits right in.

3. Good story. "Knowing how to solve problems is not the same as knowing how to teach" is a line everyone agrees with, and the paper included a table comparing it against a row of closed-source large models. Both pieces could be turned straight into an article.

4. Never stopped shipping. In the two or three weeks after the paper, something new came out every week: FP8, training code, evaluation, a pip package. Each update gave media a fresh angle and made engineering-minded people think the repo was "usable" and worth starring.

Lots of bookmarks, little use

653 stars, against official models on Hugging Face that total fewer than 10 likes and roughly 970 downloads over 30 days, and 61 PyPI downloads a week. Most people starred it because it's "from Peking University, SOTA, bookmark it for later." I couldn't check downloads on ModelScope in China; actual usage there may be somewhat higher.

The table beside it is what matters more:

At the education-model layer, the best performers reach only a few hundred to a thousand stars; DeepTutor, which became an app, has over eighty times that. Open-source education stars flow to things people can use right away, not to models.

If it were you

C: you can learn an academic open-source launch playbook, but "train another education model" isn't worth chasing. Its stars came from Peking University, the HF charts and channels like Datawhale, which followers can't access; even if you got them, the ceiling is a few hundred stars.

Who it suits: graduate students and labs publishing papers with a model or dataset in hand who want their repo to gain stars; or developers who want to build open-source education AI apps and need a ready-made open model as a base.

Who it doesn't suit: solo developers without academic backing hoping to train a niche small model for stars (compute, evaluation and the "SOTA" claim all cost money and reputation); anyone aiming for several thousand stars (that scale lives in the application layer, see DeepTutor).

How to do it:

1. Researchers copying the release playbook: grab a spot on HF Daily Papers the day the arXiv paper goes up, put the chart ranking in the Chinese article headline, get an open-source community Official Account like Datawhale to publish the press release, syndicate it to Sohu, Zhihu and community forums, then submit to AI directory sites; add a group QR code to the README. For two or three weeks after launch, ship one new thing each week (quantized versions, training code, evaluation, pip package).
2. Developers building the layer they skipped: take open education models and data like OmniEdu and build something people can use right away, such as a local tutoring tool that generates problems aligned to Chinese curriculum standards and explains mistakes, an homework diagnosis plugin for teachers, or a "tutor" Skill plugged into Claude Code or Codex. Stars and real usage come at the application layer.
3. Build it in another language: no one has yet made open-source curriculum-aligned versions for Japan, Korea or Southeast Asia. If you have connections in education in those markets, telling the same story in another country means far less competition than in Chinese.

How to get users: on the academic route, what matters most is a ranking on HF Daily Papers and a Chinese press release that multiple platforms will carry. On the application route, name "based on Peking University's OmniEdu" in the title to borrow its reputation, then pitch it to Chinese GitHub recommendation accounts and AI directory sites.

When to move: the policy heat around "education-specific large models" is still there, and media still want to cover it. Capture the application-layer opportunity while OmniEdu's momentum lasts, ideally before the spring semester starts in 2027.

Signs you've read it wrong: (1) OmniEdu climbs to two or three thousand stars within a month or two, or Hugging Face downloads clearly pick up, meaning demand at the model layer is bigger than estimated here; (2) an app someone built on it passes a thousand stars in a month, proving the application route works and you should move now; (3) conversely, if an app like DeepTutor builds curriculum alignment and error diagnosis in fully, there's less room left for new apps.