The extent to which AI can advance in the education and training sector hinges crucially on the progress of domestic-developed models.

Education and training companies all seem to be building their own AI capabilities, but what truly determines how far these applications can go is the underlying foundation model. Constrained by security, stability, and cost considerations, domestic education and training enterprises are actually more reliant on domestic models. This is exactly the significance of Kimi K3: every step forward made by domestic models can push more educational AI scenarios from experimentation to real-world implementation.

Recently, Moonshot AI launched its new-generation model, Kimi K3.

Just like previous large model releases, the outside world first focused on its parameters, benchmark scores, and technical capabilities: what level it has reached in programming, knowledge work, and complex task execution, and how large the gap remains between it and the world's most advanced models. After the launch, Kimi temporarily suspended new user subscriptions to alleviate computing power pressure.

But for the education and training industry, what deserves more attention than technical rankings is what changes the improved capabilities of domestic models represented by Kimi K3 will bring to educational AI.

I. The AI capabilities of education and training companies are not entirely under their own control

Over the past few years, many education and training companies have set up their own AI teams.

Some leading companies have their own algorithm, data, product, and engineering teams, and have accumulated massive amounts of data on exam questions, courses, student behavior, and teacher-student interactions. They can train small models tailored to educational scenarios, fine-tune general-purpose models, and use knowledge bases, prompts, and workflows to enhance AI performance on specific tasks.

But these efforts are completely different from training an advanced foundation model.

Competition in foundation models requires massive capital and top-tier talent. Even for large internet companies, participating in foundation model competition is a long-term, capital-intensive, and high-risk investment. For education and training companies that have undergone industry adjustments and must balance revenue, profits, and cash flow, undertaking such investments independently is not realistic. Therefore, the more practical choice for education and training companies is to build applications on top of external foundation models.

What education and training companies are truly good at is understanding educational business: what kind of explanations are suitable for students of different ages, what kind of feedback can help students identify problems, how to integrate AI into courses, question banks, and teaching processes, and how to judge whether an AI tutoring session is truly effective.

But these capabilities determine where AI is used and in what form, yet they cannot fully determine the maximum potential of AI itself.

When a foundation model lacks sufficient reasoning capabilities, education and training companies can add knowledge bases and standard answers, but it is difficult to turn it into an AI teacher that can stably handle complex problems. When a foundation model cannot accurately understand images, speech, and long conversations, even the most sophisticated designs at the application layer can hardly completely fill the gaps in underlying capabilities.

Therefore, the AI capabilities of education and training companies appear to be reflected at the product layer, but their foundation is built on the public capabilities provided by model enterprises. Education and training companies can decide where AI is applied, but they can hardly determine how far the fundamental capabilities of AI can progress.

II. Compared with the world's most powerful models, education and training companies are more reliant on domestic models

Since the AI capabilities of education and training companies depend on external foundation models, the next question is: do they rely more on the world's most powerful models, or on domestic models?

The world's most powerful overseas models are certainly important.

They continuously push the upper limits of capabilities in reasoning, programming, multimodality, and complex task execution, and allow domestic education and training companies to see what kinds of products may emerge in the future. Many AI product forms were first validated on advanced overseas models.

But the existence of a certain capability in the world does not mean that domestic enterprises can stably access and use it.

For education and training companies, choosing a model as their business foundation cannot only consider whether it is the most capable. Three more practical issues must also be considered: security, stability, and cost.

The first is security. Education is not an ordinary information service. It involves not only knowledge transfer, but also value guidance, minor protection, and ideological security.

The AI systems of education and training companies may process students' personal information, learning records, teacher-student conversations, and family situations; AI products for students will also directly answer questions about history, culture, society, and value judgments. Once a model is integrated into educational products, it is no longer just a tool to help employees improve efficiency, but can also become an information entry point for students to acquire knowledge and understand the world.

Therefore, for education and training companies, a model must first be safe and controllable, before we can consider whether it is sufficiently capable. This is not an additional condition, but a prerequisite for a model to be integrated into core businesses.

The second is stability. Even if an overseas model has stronger capabilities, if there are uncertainties in its access channels, service continuity, and usage policies, it can hardly become the core production foundation for domestic education and training companies.

The AI applications of education and training companies are not called occasionally. Once a model is integrated into homework grading, student Q&A, sales quality inspection, customer service, or course products, it means that a large number of businesses rely on its operation every day. Interrupted interfaces, restricted calls, or changes in service rules may affect user experience, or even cause business processes to halt. Enterprises can use less stable models for research and testing, but it is difficult to build their core businesses on a system whose long-term availability they cannot control.

There is also the cost factor. Many AI applications in the education and training industry are characterized by high-frequency and massive calls. The cost of the world's most powerful overseas models is roughly 5 to 10 times that of domestic models. Every time a student submits a question, a model call may be required; one AI Q&A session often involves multiple rounds of conversations; sales, customer service, and tutoring teachers generate large amounts of audio recordings and text every day; educational content generation, user tag extraction, and service quality inspection also require continuous batch operation.

A difference of a few cents per call may not seem significant, but when the call volume reaches hundreds of thousands or millions, the cost difference will directly determine whether the business model of a product is viable.

Therefore, what education and training companies truly need is not necessarily the world's most capable model, but a sufficiently capable model that can be used on a large scale at an affordable cost on the premise of security and stability.

This is precisely the special significance of domestic models. The world's most powerful models determine what the education and training industry can imagine, while domestic models determine what domestic education and training companies can actually deploy. Even if enterprises access multiple models at the same time, out of considerations of security, stability, and cost, domestic models are more likely to become the fundamental foundation with the widest coverage and the highest call frequency.

III. Only when domestic models cross the capability threshold can AI scenarios achieve concentrated implementation

At the beginning of 2025, the release of DeepSeek-R1 once sparked a boom in domestic AI applications.

After the release of DeepSeek, a large number of enterprises, government agencies, and products quickly announced their integration with it. Scenarios such as government services, policy Q&A, enterprise knowledge bases, intelligent customer service, official document processing, code development, and data analysis saw a surge of applications in a short period of time.

These application scenarios did not suddenly emerge after the release of DeepSeek.

Enterprises have long known that AI can be used in customer service, office work, knowledge retrieval, and data analysis, and many institutions have conducted tests before. What DeepSeek truly brought about was a model option that offers strong capabilities, low costs, high openness, and can be stably deployed by domestic enterprises.

It did not invent these scenarios, but unlocked them.

Before a model crosses the capability threshold, an AI project may already be technically operational, but not economically viable. For example, AI can generate educational research content, but a large amount of content still requires manual word-by-word inspection; it can analyze sales conversations, but the recognition results often miss key information; it can answer students' questions, but its performance on complex problems is unstable.

AI seems to have completed most of the work, but employees still need to follow up to check, modify, and rework. Enterprises not only have to pay for model costs, but also cannot reduce their original manpower input, and sometimes even need to add an additional review process. What truly determines whether an AI project can be scaled up is often not that the model "can do" a certain task for the first time, but that the model can reduce the manual fallback cost to an acceptable level for the enterprise.

In the past, every result required manual review, but later only sampling inspection was needed; previously, employees had to make extensive modifications, but later they only needed to handle a few anomalies; originally, AI could only provide references, but later it could independently complete a standardized process. When domestic models cross such thresholds, changes will not only occur in one project, but may simultaneously appear in a batch of scenarios.

The education and training industry is no exception.

Scenarios such as AI-generated exam questions, teaching plan creation, homework grading, Q&A and problem explanation, sales conversation analysis, user tag extraction, and service quality inspection were not discovered today. In the past, some of them could only stay in the proof-of-concept stage; some could only serve as auxiliary tools for employees; and some, although already launched, were difficult to truly reduce costs and improve efficiency due to excessively high manual review costs.

As domestic models continue to improve in capabilities such as reasoning, instruction following, multimodality, long-text processing, and task execution, a batch of previously unstable tasks will become stable, results that previously required extensive modification will gradually become directly usable, and high-frequency calls that were previously unaffordable may become feasible due to cost reduction.

Therefore, only when domestic models cross the threshold of production readiness can these scenarios be transformed from ideas and experiments into real projects and products in a concentrated manner.

IV. The progress of educational AI depends on how far domestic models have advanced

Let's return to Kimi K3.

Its significance to the education and training industry does not necessarily lie in the immediate announcement of integration by a certain education and training company, nor will it immediately spawn a brand-new educational product.

More importantly, it represents that the capability boundary of domestic models is still moving forward. Every improvement made by domestic models in reasoning, multimodality, long-text processing, and complex task execution may reduce the cost of using AI for education and training companies, cut down manual review and rework, and expand the scope of tasks that AI can undertake.

On the surface, the AI progress of the education and training industry comes from product launches, system upgrades, and business transformations of various companies. But at a more fundamental level, these changes are all built on the continuously improving public capabilities of domestic models.

Education and training companies can decide to prioritize AI grading, AI educational research, AI quality inspection, or AI teachers, but how far these products can go, how many users they can cover, and ultimately whether they can form a viable economic model, largely depend on how far domestic models have advanced.

The world's most powerful models tell the education and training industry what AI may be able to achieve in the future. Domestic models determine when these capabilities can truly enter the products, classrooms, and business processes of Chinese education and training companies in a large-scale, stable, and safe manner.

This article is from WeChat official account "Duozhiwang" (ID: duozhiwang), author: Tcoh, and published with authorization from 36Kr.