Inside 'ZhiHui KePu': The Rural-Education Award Winner Built with Qwen and Manim

Core Highlights

The first-place rural-education entry at the inaugural 'Small but Mighty' competition, 'ZhiHui KePu,' offers a replicable engineering pattern that automatically turns knowledge points into teaching animations. Instead of drawing frame by frame by hand, it uses a large model plus an animation engine to turn 'explaining knowledge' directly into 'an editable video students can watch,' putting the otherwise scarce ability to produce high-quality science content into the hands of ordinary teachers who may know the subject but cannot animate. The win is significant because the bottleneck in rural education is rarely the curriculum, it is the capacity to present it vividly, and a system that manufactures that capacity on demand changes the equation for under-resourced schools.

What the Capability Actually Does

'ZhiHui KePu' uses the Qwen3.5-397B-A17B large model together with the Manim animation engine. The whole system is completed by multiple agents collaborating in stages: a planning agent first decomposes the knowledge topic and sets the outline; a draft agent generates the animation script and shots; an implementation agent calls Manim to write runnable code; a review agent checks logic and visuals; and a synthesis agent finally outputs the finished film. The most practical design is the auto-repair mechanism: when a segment fails to render, the system automatically extracts the error log, locates the problem, and rewrites the code to re-render, without a human standing by to fight the fire every time something breaks during a long overnight batch. The five-stage shape is what makes the thing robust: each agent has one job, so when output is wrong you know exactly which stage to interrogate rather than staring at one monolithic black box.

Technical Details

It splits 'knowledge topic to teaching animation' into a five-stage pipeline, each stage owned by a dedicated agent, with stages passing structured intermediate artifacts, which is easy to debug and reuse because each handoff is inspectable. Manim itself is a code-driven math animation library, so every frame produced is editable and parameterizable; a teacher can change the text, colors, and pacing to match a specific class. Auto-repair relies on a log-parsing plus code-rewriting loop, turning 'render crashed' from manual firefighting into system self-healing, so the whole flow never has to start over from scratch just because of one error in a single scene. Because the deliverable is code, not a baked video, the teacher owns the master copy and can regenerate variations for different grade levels without re-prompting the model from zero.

Comparison with Competing Approaches

Compared with general text-to-video tools that only give a black-box short clip, the output of 'ZhiHui KePu' is a structured, editable engineering file, better suited to education where accuracy and iterability are demanding and a wrong formula cannot be left in a video shown to children. It also does not rely on a single model brute-forcing the task, but uses multi-agent division of labor to break the complex task apart, giving stronger stability and controllability, and making errors easier to locate and correct when they do appear in a specific stage. The contrast is between a toy that produces one unrepeatable clip and a workshop that produces a correctable asset a school can keep and improve, which is the difference between a demo and a tool.

Industry Impact and Use Cases

Simply put, this proves that 'AI plus animation engine plus multi-agent' can genuinely solve the science-popularization teaching problem under rural teacher shortages. The same engineering pattern can also migrate to tracks beyond science popularization, such as experiment demos, policy explainers, and skill training, letting high-quality visual teaching content be produced in batches at low cost, and spreading premium educational resources from the hands of a few to a much broader grassroots base that has long lacked access to such material. The deeper lesson is methodological: the competition entry is less a product than a recipe, and a recipe is something any motivated teacher or county education office can now replicate with commodity models, which is how a one-off win turns into a durable public good. The five-stage pipeline is also a teaching instrument in its own right, because it makes the production of an explanation visible: a student or a fellow teacher can open the intermediate artifacts and see how a topic was decomposed, where the logic was checked, and what had to be fixed, which is a form of metacognition that a finished video alone never offers. Once the barrier to producing high-quality visual explanation drops, the scarce resource stops being the ability to animate and becomes the ability to choose what is worth explaining, a genuinely human judgment that models cannot make for us. In that sense the project does not replace the teacher, it returns the teacher to the part of the job that matters most, deciding what their students ought to understand and why it is worth their attention.