Zhihui Science Technical Breakdown: Qwen and Manim Challenge the One-Click AI Video Model

Zhihui Science won first prize in a rural education track by turning lesson topics into editable animations through a five-stage AI workflow. This Zhihui Science technical breakdown examines the reported system behind that result, including Qwen3.5, Manim, specialized agents, and automated code repair.

The project emerged from the first Xiaoyou Kewei AI Open-Source Public Welfare Innovation Challenge. The competition brought together nearly 800 teams and received 455 entries across four public-interest categories. Zhihui Science competed in the rural classroom AI teaching-assistant category.

Its important opponent is not another contest entry. It is the familiar one-click generation model, where a teacher submits a prompt and receives an opaque video with limited control. Zhihui Science takes a different route by generating plans, scripts, code, review decisions, and editable scene files.

That choice makes the project more than another demonstration of AI-generated media. It tests whether structured agent workflows can make educational content easier to inspect, revise, and maintain.

What Zhihui Science Actually Changed

Zhihui Science reframes AI video generation as a supervised production pipeline instead of a single creative request.

The project was built by the Tashan Interdisciplinary Innovation Association at the University of Chinese Academy of Sciences. It won first prize in the competition's rural education category, according to the available event reporting.

The challenge began in September 2025 and held its final in Hangzhou on December 29. A reported 42 finalists reached the final evaluation, while 16 projects received awards across four public-interest tracks.

Those details matter because Zhihui Science was evaluated as a response to a defined classroom problem. It was not presented only as an experimental animation generator.

The contest focused on rural education, services for older adults, support for autistic children, and accessibility. Other winning projects included an AI memoir tool, a children's picture-book project, and open-source assistive glasses.

A competition report described Zhihui Science as a bridge between distant digital resources and rural classrooms. That framing identifies the project's intended users: teachers who need instructional materials but lack dedicated animation teams.

A later technical post from the Qwen organization reportedly detailed the architecture. Its account centers on Qwen3.5-397B-A17B, the Manim animation engine, and several cooperating agents.

The accessible summary describes five production stages: planning, drafting, implementation, review, and composition. Each stage converts the previous stage's output into a more concrete artifact.

The process begins with a knowledge topic rather than a complete storyboard. A planning agent interprets the topic, identifies concepts, and divides the explanation into scenes.

A drafting stage then turns that plan into a visual narrative. It decides what should appear, which relationships need emphasis, and how each scene should progress.

The implementation stage produces Manim code. Manim is a Python animation engine that represents visual elements and motion as programmable objects.

A review stage evaluates the generated result or its underlying specifications. The composition stage then assembles accepted scenes into the final instructional video.

This separation is the central change. A one-step generator hides most decisions inside a model response. Zhihui Science reportedly exposes intermediate plans and code that humans can inspect.

The project also adds a repair loop around rendering. When Manim fails, the system extracts error information and sends it back for diagnosis and correction.

The final output therefore includes more than a video file. It includes a structured path from topic to scenes, plus code that can be edited and rendered again.

That distinction creates the article's central tension. Greater control requires more engineering, but educational media often needs exactly that control.

Why Rural Teachers Put the Workflow Under Pressure

A classroom animation system succeeds only when teachers can correct it without rebuilding the entire lesson.

Rural education gives the project a socially compelling setting, but it also creates demanding product constraints. Schools can differ widely in staffing, connectivity, devices, and access to technical support.

A polished demonstration does not resolve those differences. The system must eventually work across ordinary lessons, limited preparation time, and inconsistent infrastructure.

The immediate pressure falls on conventional content-production workflows. Teachers often choose between static materials, generic online videos, and custom resources that take substantial time to produce.

Generic videos can explain the right subject while missing a class's precise learning objective. They may use unfamiliar notation, introduce concepts in the wrong order, or assume knowledge that students lack.

Custom animation offers more control, but writing animation code requires specialized skills. Manual video editing creates another set of costs involving assets, timelines, narration, and revision.

Text-to-video systems reduce some production work. However, they usually optimize for visual plausibility instead of exact symbolic relationships.

That tradeoff becomes obvious in mathematics and science. A diagram must preserve labels, geometry, sequence, scale, and causal relationships across frames.

A visually attractive mistake can be more damaging than a plain slide. Students may remember the motion while absorbing an incorrect relationship.

Zhihui Science responds by treating the teacher's request as a specification. The agents transform that specification into intermediate documents before generating the final animation.

This approach does not guarantee accuracy. It does create several points where a teacher, reviewer, or future validation system can intervene.

The planning artifact can reveal a missing concept. The scene draft can expose a confusing explanation. The code can reveal an incorrect formula or visual transformation.

That makes editability more than a convenience. It becomes a method for controlling instructional risk.

The project also pressures one-click educational media tools because its output remains programmable. A teacher or developer can change a label, duration, color, formula, or sequence without regenerating every scene.

Programmatic output can also support reuse. A reviewed scene template could be adapted across related lessons while preserving approved visual logic.

This benefit still depends on the interface. Most teachers will not want to debug Python or inspect raw scene classes.

A successful deployment must hide unnecessary complexity while preserving access to meaningful controls. The difficult design question is where that boundary should sit.

Teachers might edit learning objectives, terminology, scene order, and pacing through forms. Technical staff could retain access to generated code and rendering logs.

Schools with limited connectivity create another constraint. The reported use of a 397-billion-parameter model suggests that generation probably requires hosted inference or substantial remote infrastructure.

The model activates 17 billion parameters for each token rather than using every parameter simultaneously. That reduces computation relative to a dense model of the same total size, but it does not make deployment trivial.

The larger point is that Zhihui Science does not remove production infrastructure. It reorganizes that infrastructure around a teacher-facing workflow.

That is still meaningful. Educational teams already manage source documents, lesson plans, examples, assessments, and review comments.

A structured knowledge workflow can keep those materials connected. Tools for a searchable knowledge base offer a related pattern for preserving sources and revisions.

The competitive pressure is therefore long term. AI video products will need to support provenance, correction, and reuse rather than treating generation as the final step.

Zhihui Science Technical Breakdown: Five Agents Around One Renderer

The project's key mechanism is the division of creative work into bounded stages with machine-checkable outputs.

The available description uses the term multi-agent system. An agent here is a model-driven software component assigned a specific task, tools, and expected output.

This does not necessarily mean five separate models. Multiple agents can use the same underlying model with different prompts, tools, context, and validation rules.

Qwen3.5-397B-A17B reportedly supplies the language and reasoning layer. Its official model card describes 397 billion total parameters, with 17 billion activated during inference.

The model uses a mixture-of-experts architecture. This design routes each token through selected expert networks instead of activating the complete parameter set.

The model card also identifies 512 experts, with ten routed experts and one shared expert activated. Its native context window is listed as 262,144 tokens.

Those specifications can support long production contexts containing source material, planning notes, scene descriptions, code, and error logs. They do not independently establish the educational accuracy of the generated content.

The first reported agent handles planning. It converts a broad topic into a sequence of learning goals and visual units.

Consider a lesson about projectile motion. A useful plan must decide whether to begin with velocity components, coordinate axes, equations, or a physical example.

That decision affects every later scene. If the plan introduces formulas before defining the visual frame, better animation quality will not repair the teaching sequence.

The second agent produces a draft. This stage translates the educational plan into visual actions, narration concepts, and scene-level structure.

The distinction between planning and drafting resembles a separation between curriculum intent and media execution. One defines what students should understand, while the other decides how to show it.

The third agent implements the scenes in Manim. According to the official Manim quickstart, scripts organize animations inside Scene classes and their construct methods.

Visual elements become mathematical objects, commonly called mobjects. A developer can create shapes, equations, labels, graphs, and transformations through code.

This representation brings determinism to parts of the workflow. A circle defined in code remains a circle, and a formula can be positioned using explicit coordinates.

It also makes outputs editable. Changing the duration of an animation or the location of a label does not require reconstructing a complete video through prompts.

Manim then serves as an execution environment and an early validation layer. Code either runs or produces an error that the surrounding system can capture.

That distinction is important because large language models can generate code that looks plausible without executing correctly. Syntax errors, missing imports, invalid object references, and incompatible calls remain common failure modes.

Zhihui Science reportedly addresses these failures through automated repair. The system collects rendering logs, identifies the likely failure, revises the code, and tries again.

This loop converts a passive model response into an iterative software process. The model does not merely answer once; it receives feedback from an external tool.

A rendering error is also more actionable than a general request to improve quality. It can identify the file, operation, or line involved in the failure.

Automated repair therefore handles a narrow but useful class of problems. It can correct execution failures without asking a teacher to interpret a stack trace.

The fourth stage reviews the output. The public summary does not provide enough detail to determine whether this review inspects code, rendered frames, educational logic, or all three.

That uncertainty matters. A successful render only shows that the animation executed. It does not show that the explanation is scientifically correct.

A review agent could compare the script against source material, inspect required concepts, or use vision capabilities to analyze rendered frames. Each method catches different errors.

Text review can verify formulas and narration. Code review can identify implementation problems. Visual review can detect collisions, illegible text, and confusing motion.

The fifth stage composes approved scenes. Composition can manage ordering, transitions, audio, and the final rendered package.

Keeping composition separate supports partial regeneration. A failed or weak scene can be revised without discarding every accepted scene.

That modularity is the strongest engineering idea in the Zhihui Science technical breakdown. Each agent produces an artifact that the next stage can consume and evaluate.

It also creates an audit trail. Teams can retain the initial request, plan, draft, generated code, review result, repair history, and final output.

Such records become valuable when educators question a claim or request a change. Developers can trace the error to a specific production stage.

The system resembles a continuous integration pipeline for media. Code generation, execution, inspection, correction, and assembly happen as connected operations.

However, the analogy has limits. Software tests can assert exact outputs, while educational clarity and conceptual understanding are harder to measure automatically.

The architecture is therefore strongest when it combines deterministic validation with human review. Rendering tests handle execution, while educators judge pedagogy and suitability.

A Successful Render Is Not a Correct Lesson

Automated repair can fix broken code, but it cannot certify the truth or teaching quality of an animation.

The available reporting confirms the competition result and describes the architecture. It does not provide a public benchmark covering accuracy, teacher workload, student outcomes, or deployment reliability.

That verification gap should shape how the project is evaluated. Zhihui Science is a promising engineering case, not established proof that agent-generated animations improve learning.

The clearest risk is factual error. A model might generate a valid Manim scene that presents the wrong equation, reverses a relationship, or omits a necessary condition.

The renderer will accept that scene. Automated repair will not intervene because the code works exactly as written.

Review agents can reduce this risk, but they may share the generator's blind spots. Agents built on the same model can confidently reinforce the same mistaken interpretation.

External grounding is therefore essential. The workflow should tie claims, formulas, definitions, and examples to approved textbooks or teacher-provided reference materials.

A teacher should also see those sources during review. Provenance cannot remain hidden inside a retrieval service or system prompt.

Visual quality creates a second risk. Programmatic animation offers precision, but generated layouts can still become crowded, poorly paced, or difficult to read.

A formula may be correct yet disappear too quickly. A moving label may overlap another object. Colors may lack sufficient contrast on classroom projectors.

These are not minor cosmetic defects. Presentation choices affect whether students can follow the reasoning represented on screen.

Localization adds another challenge. Rural classrooms do not form a single audience, and educational content may need regional terminology, local examples, or language support.

A model trained across broad internet data may default to examples that feel distant from students' daily experience. It may also produce inconsistent translations of technical terms.

Privacy deserves attention as well. The reported workflow begins with knowledge topics, which presents less risk than systems processing identifiable student records.

However, future versions might accept student questions, assessment data, classroom recordings, or teacher documents. Those inputs would require clear retention and access policies.

UNESCO's education guidance calls for human-centered validation, privacy protection, and age-appropriate use of generative AI.

That guidance fits the project's most credible deployment path. Teachers should remain accountable for the lesson, while agents accelerate preparation and revision.

Infrastructure remains a practical concern. The flagship Qwen model is large even though its mixture-of-experts design activates only part of the network.

Remote inference introduces dependency on connectivity, service availability, and data-transfer policies. Local inference demands hardware and operational expertise that many schools will not have.

A project aimed at underserved classrooms must therefore distinguish content production from classroom playback. Teachers might generate materials where connectivity exists, then export videos and source packages for offline use.

The source package is especially important. A plain MP4 can play offline, but it cannot be easily corrected when a teacher finds a problem.

Distributing editable projects creates its own maintenance burden. Manim versions, fonts, media assets, and software dependencies can change over time.

The team will need reproducible environments, versioned templates, and dependable export formats. Otherwise, yesterday's editable lesson can become tomorrow's broken project.

There is also no disclosed evidence that the five-stage workflow consistently outperforms a simpler model-plus-template system. More agents increase opportunities for correction, but they also add latency and failure points.

A planning agent can misunderstand the topic. A drafting agent can distort the plan. An implementation agent can simplify the draft incorrectly.

Review agents can generate false alarms or approve subtle mistakes. Repair loops can fix one defect while introducing another.

The architecture needs measurements at each boundary. Teams should track plan acceptance, first-render success, repair success, factual corrections, manual editing time, and final teacher approval.

Student outcomes require separate evaluation. Faster content creation does not automatically produce stronger understanding or retention.

A credible pilot would compare lessons using identical objectives and similar teachers. It would then measure preparation time, error rates, comprehension, and delayed recall.

None of these limitations invalidate the project. They define the work required to move from a contest-winning prototype to a dependable educational system.

The key is to avoid confusing controllability with correctness. Editable code makes correction possible, but qualified reviewers and validated sources make correction meaningful.

Three Signals That Will Show Whether the Model Travels

The next test is whether Zhihui Science can turn a convincing architecture into repeatable classroom production.

The first signal is a public implementation package. That could include source code, agent prompts, schemas, example projects, evaluation scripts, and deployment instructions.

The competition emphasized open-source public-interest innovation. A usable release would let independent developers inspect how responsibilities move between agents.

It would also clarify which components are reusable. Teams could determine whether the repair loop works only with Manim or supports other code-driven media tools.

A public package would strengthen the central claim by showing that the system is an engineering pattern rather than a staged demonstration.

An incomplete release would weaken that claim. Screenshots and final videos cannot reveal how frequently generation fails or how much manual intervention occurred.

The second signal is evidence from sustained teacher use. The most useful figures would cover completed lessons, repeat users, editing time, rejection rates, and subjects supported.

A one-time workshop can show that teachers understand the concept. Repeated use shows whether the workflow fits lesson preparation under ordinary deadlines.

Teacher edits deserve special attention. If educators repeatedly rewrite plans but rarely touch generated code, the product should prioritize planning controls.

If code repairs dominate the workflow, the team may need stricter templates or constrained generation. If factual corrections dominate, stronger grounding should come first.

Usage evidence should also include offline and low-bandwidth settings. A rural education system cannot rely only on demonstrations conducted with stable cloud access.

The third signal is an independent educational quality evaluation. Reviewers should assess factual accuracy, curriculum alignment, readability, pacing, accessibility, and age suitability.

That evaluation should inspect both accepted outputs and rejected attempts. Looking only at published examples would hide the system's true error distribution.

Independent review would strengthen the project if teachers approve most outputs after limited revision. It would weaken the claim if specialist intervention remains extensive.

These signals matter beyond one competition. The five-stage pattern can travel to scientific visualization, public health explainers, accessibility materials, and other structured media.

Its portability comes from the interfaces between stages. Planning, drafting, implementation, review, and composition are general production functions.

The specific tools can change. Another workflow might replace Manim with a diagram renderer, slide framework, simulation engine, or web component library.

The same repair principle can also travel. When an AI system operates a deterministic tool, execution feedback can guide another attempt.

However, each new domain needs its own definition of correctness. A code repair loop cannot replace medical review, accessibility testing, or curriculum validation.

Zhihui Science is most interesting when viewed through that constraint. It does not solve educational content generation with one larger model.

Instead, it places a large model inside a process where plans, code, logs, and revisions become visible artifacts.

That is a more realistic direction for AI-assisted knowledge work. The model performs several bounded tasks while external tools and human reviewers supply feedback.

For developers, the immediate lesson is to design intermediate artifacts before adding more agents. Each stage needs a clear input, output, validation method, and recovery path.

For education buyers, the right questions concern control and evidence. Ask who approves content, what remains editable, which sources ground claims, and how failures are recorded.

For teachers, the decisive measure is simpler. The system should reduce preparation work without transferring technical troubleshooting into the classroom.

The Zhihui Science technical breakdown therefore ends with three concrete tests: an inspectable release, repeated teacher adoption, and independent educational review.

Until those signals arrive, the project should be treated as a well-structured prototype with an important engineering thesis. Editable generation can be safer than opaque generation, but only when schools can verify what the system creates.

The next move belongs to educators and implementers. Which matters more in your setting: faster first drafts, clearer source tracking, or easier correction after generation? Start with that constraint, then evaluate whether a staged agent workflow genuinely reduces the work around it.