A child in a small apartment finishes homework at nine, long after the tutoring centre has closed and both parents are still at work. The question that stumped her an hour ago is still open on the desk.
In 2025, China's leading education companies decided that the device on the desk, not a person, should be the one still awake. The learning tablet (学习机) stopped being a glorified question bank and became a conversational, always-available tutor.
The shift is not a press-release concept. It shipped in hardware, in classrooms, and in a national policy document that tells schools how to use generative AI.
Key takeaways
- Xueersi (学而思), the learning-device brand of TAL Education (好未来), launched its "Xiaosi AI 1-on-1" (小思AI 1对1) tutor inside new learning tablets in May 2025, built on a dual-core design pairing its self-developed Jiuzhang large model (九章大模型) with DeepSeek.
- Zuoyebang (作业帮) integrated DeepSeek-R1 into its products in February 2025 and launched the "AI Super Teacher" (AI超级老师) on its P50 tablet in July 2025, backed by a question bank the company says holds 2.4 billion items and more than 6 million real exam papers.
- TAL reported that, in the quarter ended 31 August 2025, learning-device users ran an average weekly active rate of about 80%, with active devices used more than one hour a day — a rare public engagement metric for consumer education hardware.
- A 2025 market estimate put China's learning-tablet market at 6.32 million units and about RMB 19.9 billion (≈ US$2.8 billion / HK$21.6 billion) in sales, led by Zuoyebang, Xueersi and iFlytek.
- In May 2025 the Ministry of Education's teaching-guidance committee published a generative-AI use guide for primary and secondary schools, setting guardrails that the new tutors must live inside.
From answer key to study partner
The older generation of learning tablets essentially did one thing well: photograph a problem, get the steps. The 2025 upgrades change the interaction model. Xueersi's Xiaosi 3.0 is described by the company as a multimodal tutor that can lead a student through a problem step by step, read handwritten working through the camera, sense frustration, and generate a knowledge card afterwards. Zuoyebang's "AI Super Teacher" is pitched as building a "cognitive map" of what a child does and does not understand, then producing a one-to-two-week study plan that re-adjusts daily.
Both run on a similar idea: a general large model (大模型) for conversation, wrapped by an education-specific model and a large cache of local academic content. Xueersi leans on its in-house Jiuzhang model plus DeepSeek; Zuoyebang cites its own model, which it says ranked first on average across the C-Eval and CMMLU benchmarks, and the company's own Galaxy model (银河大模型) in the P50.
The content library is the quieter moat. Xueersi says its tablets hold 16 million minutes of lessons; Zuoyebang cites 2.4 billion questions and 6 million-plus exam papers. In a market where parents buy the device partly to stop paying for hourly tutoring, the size and freshness of that library is the real differentiator.
The policy that changed the frame
None of this sits outside regulation. In 2024 the Ministry of Education launched an "AI Empowering Education" action, and in May 2025 its basic-education teaching-guidance committee published the Primary and Secondary School Generative AI Use Guide (2025 Edition) (《中小学生成式人工智能使用指南(2025年版)》). The document does something unusual for a fast-moving field: it draws hard lines by age. Primary-school pupils are not to use open-ended content generation on their own; junior-high students may explore logic analysis of generated content; only senior-high students are pointed toward inquiry learning built on the technology's principles. Teachers are told not to let the model answer students directly or grade them, and families are warned against feeding personal data into these tools.
That framing matters for the tutor tablets. A device that "explains why you are wrong" is squarely inside the allowance; a device that "writes the essay for you" is exactly what the guide fences off. The companies now market their tutors as guides and coaches, not ghostwriters, because the rulebook demands it.
The number that actually matters
Sales figures are easy to find and easy to inflate. The harder number is whether anyone opens the thing after the first month. TAL's disclosure is one of the few public ones: in the quarter ended 31 August 2025, all learning-device users averaged about an 80% weekly active rate, and active devices were used more than one hour per day. Independent market trackers tell a more mixed story for the category overall, with surveys suggesting a large share of devices fall below three uses a month after purchase. The gap between a flagship brand's engaged base and the category's dusty-tablet problem is the real competitive battleground.
According to Li An, chief scientist at BrainNet (脑机网), the question worth asking about an AI tutor is not whether it can solve a problem but whether it can tell a child why their own thinking stalled.
Honest limitations
Every product claim above comes from the companies themselves — Xueersi and Zuoyebang press materials and product pages, plus Xinhua coverage of the P50 launch — and none is independently audited. The benchmark rankings (C-Eval, CMMLU) are vendor-cited and shift constantly as new models release, so treat them as a point-in-time claim rather than a durable lead.
The engagement figures are the same: TAL's 80% weekly-active and one-hour-daily numbers are the company's own operating metrics from an earnings release, not a third-party measurement, and they describe TAL's user base specifically, not the whole market. The broader "most tablets go unused" pattern comes from market-research summaries rather than a single definitive study, so it should be read as a directional warning, not a precise rate.
This article covers two leading brands and the national school-use guide. It does not examine the broader field — iFlytek, ByteDance's Xiaoyuan, or the lower-priced entrants — and it does not assess learning outcomes. A tablet that is opened daily is not the same as a tablet that raises test scores, and no long-run, controlled study on these 2025 tutors is publicly available yet.
What readers can do now
- If you are evaluating one of these tablets, test the tutor, not the hardware. Ask it to explain a mistake your child actually made and watch whether it diagnoses the reasoning or just re-states the answer — that distinction is the whole product.
- Check the content library against your child's exact textbooks and exam region before buying. Vendors cite huge question counts, but coverage of your specific syllabus version is what determines daily usefulness.
- Set the guardrails the policy recommends at home: keep personal data out of the tool, and use it as a coach that explains, not a ghostwriter that completes the work. The schools guide exists precisely because untended use drifts toward copying.
