A teacher in a provincial Chinese classroom drops a stack of math papers into a desktop scanner the size of a printer. Forty seconds later the pages slide back out — red marks, a step-by-step correction, and a typed note on why the student lost points, already printed in the margin. No human hand touched the grading.
That machine is iFlytek's Spark intelligent grader (星火智能批阅机), and it has become a quiet fixture in more than 2,000 schools across China. It is one of the clearest examples of a large model (大模型) leaving the cloud and landing on a teacher's desk.
The machine that grades like a senior teacher
The grader is built on iFlytek's Spark (星火) large model, the same family of models that powers the company's AI learning machines (学习机) and voice products. Where most auto-grading stops at ticking right or wrong, this device tries to explain.
It reads handwritten and printed work across subjects, and for math and physics it scores multi-step solutions the way a human would — awarding partial credit, flagging the exact step where a student went wrong. For Chinese and English essays it checks vocabulary, grammar, and structure. iFlytek reports that its grading accuracy tops 97% and that the device cuts a teacher's marking time by roughly 94%.
The part that matters for learning is not the red pen. The grader organises every mistake into a three-tier system — knowledge mastery, thinking method, and work habit — spread across more than 4,000 fine-grained error labels. A wrong answer about fractions is no longer just "wrong"; the system can say the gap is conceptual, procedural, or careless.
The same diagnostic logic follows the student home. iFlytek's consumer AI learning machine (学习机) pairs the Spark engine with DeepSeek and acts as a one-on-one tutor: it walks a child through a stuck problem, then recommends the next exercise based on the weakness the model just diagnosed. School and home share one idea — capture the mistake, name the cause, act on it.
From six million pages to one student's blind spot
Scale is where the story gets interesting. According to iFlytek, the grader family now processes over 6 million homework pages a day nationwide. That volume is impossible for any human grading pool, and it is the raw material that turns a gadget into a diagnostic instrument.
After scanning, the machine prints a class report and a lecture note for the teacher, and pushes a personal correction plan with variant exercises to each student. The loop is meant to close the gap between "the test is graded" and "the student actually improves."
iFlytek first shipped the grader in June 2024 and has since extended it to 2,000-plus schools for daily homework and exam use. The broader smart-education business reaches further: the company says its products sit in more than 60,000 schools and have served over 160 million teachers and students across 33 provincial-level regions in China.
A district that cut busywork by more than a third
Concrete district data is rarer than marketing claims, so one number from Shenyang is worth a close look. In the Dadong district, where iFlytek tools have run for three years, the company reports that inefficient, repetitive student practice fell 36.4% from the 2022 baseline, while average knowledge-point score rates rose 51.7% by the end of 2025. The district has since stood up more than 150 school- and district-level agents.
Numbers like these come from the vendor and from a single district, so they should be read as a deployment signal rather than proof. But the direction is consistent with what teachers describe: the machine absorbs the mechanical load, and the teacher spends the freed time on the parts of teaching a model cannot do.
Why a Chinese ed-tech story travels
iFlytek (科大讯飞) is a Hefei-based, publicly listed company that has worked on speech and language AI since 1999. Its education arm is now large enough to matter financially — smart-education revenue reached about 5.76 billion yuan (≈ US$811M / HK$6.34B) in the first three quarters of 2025, roughly a third of the company's total.
For global readers, the takeaway is not "China has an app." It is that a national education system is quietly routing millions of daily grading decisions through a model, and the model's job is to hand the insight back to a human, not replace one. Exam-driven, teacher-scarce systems make natural laboratories for this pattern, and similar "capture, diagnose, advise, keep the teacher in charge" loops are now appearing in healthcare and government service across the country.
Honest limitations
Every figure above about grading volume, accuracy, and the Dadong district comes from iFlytek's own disclosures and a vendor-run regional case; independent academic audits of learning outcomes are not yet public. The 60,000-school and 160-million-user reach are company-reported cumulative counts, not a measure of daily active use. This article does not assess data-privacy practices, nor compare iFlytek's model against overseas tutoring systems head to head.
What readers can do now
- If you build or buy education software, ask vendors for the error taxonomy, not just the accuracy score — a 97% grade means little without a reason attached.
- Teachers evaluating auto-grading should pilot it on one subject for a term and track whether marking time actually converts into more feedback time.
- Watch for independent studies: the real test of tools like this is whether student scores move once the busywork is removed, not whether the machine marks faster.