A parent watches their child stare at a math problem for ten minutes, then snap the photo and wait. In the past, the app would return an answer copied from a fixed answer key. Now the same child gets a question back: "What step do you think comes next, and why?"
That shift — from giving answers to coaching thinking — is what TAL Education (好未来) built its MathGPT (九章大模型) to do. It is one of the clearest examples of a homegrown education large model (大模型) moving from a press release into a product students actually hold.
Key takeaways
- MathGPT (九章大模型) is TAL Education's education-focused large model (大模型), filed and approved by Chinese regulators in November 2023 as one of the first batches of compliant education models.
- TAL describes it as the first education-domain model certified by Chinese authorities, trained on the company's two decades of teaching content and knowledge graphs.
- Its main consumer surface is the Xueersi (学而思) learning device, where features like "AI dialogue study", an essay grader and a "ask-anytime" tutor run on the model.
- The company reports the model has been deployed in more than 1,000 schools across over ten provinces, including Beijing, Shandong and Hubei.
- TAL is a New York-listed company (NYSE: TAL), which makes its AI education push one of the few with a public, audited parent.
From answer key to coach
The core idea behind MathGPT (九章大模型) is not speed but diagnosis. Traditional adaptive learning judged a student by which problem they got wrong. A large model (大模型) can instead break a single problem into steps, ask about each step, and locate the precise misconception — a weak calculation habit versus a gap in logic.
On the Xueersi (学而思) device, "AI dialogue study" walks a student through that step-by-step exchange. A separate essay feature drafts word-level and paragraph-level feedback in Chinese and English, then suggests structure and offers one-click polishing. The "ask-anytime" tutor, MathGPT Ask Anytime (九章随时问), runs as its own app and mini-program, behaving less like a search box and more like a patient one-on-one teacher.
TAL has claimed, in its own benchmarking, that the model topped the MathEval leaderboard in March 2024 and outperformed GPT-4o on a set of gaokao (college entrance exam) math problems. Those are the company's figures and should be read as marketing-grade, not independent proof.
Why an education company built its own model
TAL could have bolted a general chatbot onto its products. It chose to train a vertical model because the failure modes of education are unforgiving: a wrong hint at the wrong moment can cement a mistake. A model tuned on curriculum, not just the open internet, is less likely to wander.
The company frames the build as breaking what it calls the "impossible triangle" of education — personalization, quality and scale, which had historically been hard to deliver together. Whether that claim holds at scale is still being tested in classrooms.
Beyond the living room
The model is not only inside devices. TAL's "智慧教育" open platform, built on the same technology, supplies adaptive homework, after-school services and learning diagnostics to schools. Reportedly more than 1,000 schools across over ten provinces use these tools, extending the AI tutor from a purchased gadget to a shared school resource.
For teachers, a companion toolset (老师帮) handles lesson preparation, question banks and parent communication — the unglamorous back office of education where AI may save more time than any headline feature.
TAL is not alone in chasing AI tutoring, but its scale and listed-company transparency make it a useful bellwether. When a NYSE-listed education firm stakes its AI story on a self-trained model rather than a generic chatbot, it signals where the industry thinks durable value lies: in curriculum-owned vertical models, not in reselling someone else's API. That choice is expensive, but it keeps the pedagogical judgment inside the company instead of leased from a general-model provider.
Honest limitations
Every performance claim here about MathGPT (九章大模型) traces back to TAL's own announcements and tests; there is no independent academic benchmark confirming the gaokao or MathEval results. "First certified education model" is the company's description of a regulatory filing, not a third-party ranking.
A learning device is only as good as the student's habit of using it, and AI tutoring can mask gaps if children outsource thinking to the model. Deployment scale in schools is reported by the vendor, not audited. The model also inherits the limits of any large model (大模型): it can confidently explain a wrong method, and it should sit alongside, not replace, a human teacher.
What readers can do now
- If you have a Xueersi (学而思) device or app, try the "ask-anytime" tutor on one hard problem and compare its step-by-step hint with the textbook method.
- Use AI tutoring as a sparring partner, not an answer machine — insist the model explain the next step before revealing the result.
- Parents evaluating any "AI tutor" should ask which curriculum it was trained on and whether a teacher remains in the loop, not just how fast it answers.