A parent hands a child a tablet after dinner. The math worksheet goes under the camera, and within seconds the device has marked every sum, circled the wrong ones, and started explaining why.
Some of those marks are right. Some are confidently wrong, because the model read a smudged digit as something else and never told anyone.
This is the daily reality of at-home tutoring machines in Chinese homes: faster than a parent, cheaper than a tutor, and not yet trustworthy on its own. The question is no longer whether the technology exists, but where it can be left alone and where a human must still check.
The model behind the device
The engine is Jiuzhang (九章大模型), also known as MathGPT, built by TAL (好未来 / Xueersi). It was among the first batch of education-focused large models (大模型) to clear China's generative-AI service filing (备案), the registry run by the Cyberspace Administration of China. It also received the highest 4+ rating in the education large-model evaluation run by the China Academy of Information and Communications Technology (CAICT).
That pedigree matters less to a family than what happens at the kitchen table. TAL ships the model inside learning devices (学习机) that pair Jiuzhang with the open-source model DeepSeek in a dual-core setup, and inside a one-on-one tutoring agent called "Xiao Si AI 1-on-1" (小思 AI 1对1), which also earned the CAICT 4+ rating and a case-certification from a Ministry of Industry and Information Technology culture center.
What the device actually does
The loop most families meet looks like this:
- Scan and grade. A photo of the page becomes a marked sheet; right and wrong answers are separated in seconds.
- Explain the mistake. Instead of just showing the answer, the model walks through the steps and points at the one that broke.
- Drill the weak spot. It generates similar problems at varying difficulty, so a child practices the exact gap rather than the whole chapter.
- Listen, not just read. Newer versions judge spoken answers and spoken reading, nudging pronunciation and confidence.
TAL says its content library behind the device runs to 16 million minutes of lessons, with 850,000 sets of past exam papers and a 700-million-question bank. Those are company figures, useful for scale but not independently verified.
Where it is already in the classroom
The home device is only half the story. A separate model, "Shi Cheng Wan Xiang" (师承万象), was built with Beijing Normal University and is described as the first basic-education large model aligned to China's new curriculum standards (新课标). It entered the Beijing Education Commission's AI application marketplace and has been trialed in more than 60 schools across the Haidian and Dongcheng districts.
That is the deployment worth watching: a model vetted by a university and a municipal education authority, placed into real classrooms rather than sold to individual parents. The schools, not the marketing, are the harder judge.
The honest break point
Here is where the marketing and the kitchen table diverge. Reviewers and parents report a clear pattern: the model is strong on structured, single-answer work — arithmetic, fill-in-the-blank, multiple choice — and shakier on anything that demands reading the physical page well.
- A child writes "155" and the model reads "155" as something else, then marks the whole solution wrong.
- A graph, a word problem, or a scraped-out correction defeats the camera before the model ever thinks.
- One parent's "it marked everything wrong" and another's "accuracy is high" often describe the same device on the same day, with nothing changed but the question type.
The lesson is simple: the failure is usually in recognition, not reasoning. The model often knows the math; it just misread the handwriting.
What this means for trust
The practical rule families are settling on maps neatly onto the product's own limits:
- Hand it the boring, structured stuff. Calculation checks, single-answer grading, and building a wrong-answer list are where the device reliably saves time.
- Always复核 the fuzzy stuff. Graph problems, word problems, and anything open-ended should be re-read by a human before a child is corrected or praised.
- Treat the explanation as a hint, not scripture. When the model's method disagrees with the school's, the school wins.
No model should hold the final say over whether a child "gets it." The score it returns is the score of its camera and its recognition, not proof of the child's understanding.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), the real test of an education model is whether it changes what a teacher does at nine in the morning, not the score it posts on a benchmark.
The signal to track is therefore not which device sells best, but which model actually changes classroom workflow — homework triage, lesson prep, and parent communication — in a way a school can audit.
Honest limitations
Jiuzhang's first-batch filing (备案) and CAICT 4+ rating are company-stated and corroborated by Chinese media, but we did not locate an independent international audit of the grading accuracy. The 16-million-minute, 850,000-paper and 700-million-question figures are TAL's own. The "60+ schools" trial refers to the Beijing Normal University co-built model in specific Beijing districts, not a national rollout. Consumer reports of misgraded handwritten work are anecdotal, drawn from reviews rather than a controlled study, but they describe a real and consistent failure mode — recognition error, not reasoning error. We did not verify per-district learning outcomes or any test-score improvement attributable to the device.
What readers can do now
- If you are a parent, use the device for grading structured work and building a wrong-answer list, but re-check graph, word, and open-ended problems yourself before correcting your child.
- If you are a school buyer, ask vendors for classroom-trial evidence from a comparable district, not consumer-app store ratings, and define where a human must sign off.
- If you build education AI, prioritize the camera and handwriting layer — most "it got the math wrong" complaints are actually "it read the page wrong."