Three ways for AI toys to survive: renting models, semi-in-house development, and building brains — who will get it right first?
At the China Toy Expo, Mituan Shuzhi, Miaoyin Habu and JD.com respectively represented three paths for AI toys: calling third-party large models to create original IP, using "Doubao + self-developed education model" to make an AI study companion, and providing hardware manufacturers with a "brain" through the JoyInside solution. In 2024, China's AI toy market was worth about 24.6 billion yuan and is expected to grow to 29 billion yuan in 2025, but the industry's return rate is concentrated between 30% and 50%. JoyInside has connected with more than 40 leading hardware brands, and connected products have seen average dialogue turns increase by more than 120% and response latency drop below 2 seconds.
Body · AI translation
The AI wave is sweeping through the toy market. Data released by the Ministry of Industry and Information Technology in November 2025 shows that the size of China's AI toy market was about 24.6 billion yuan in 2024 and is expected to grow to 29 billion yuan for the full year of 2025. Some research institutions predict that the size of China's AI toy market will climb to 34.2 billion yuan in 2026. Data from the China Toy and Juvenile Products Association shows that the share of products equipped with generative AI dialogue functions in China's smart toy market rose from 12% in 2023 to 34% in 2025. The market is rising, and so are the risks. A number of media outlets and research institutions have analyzed the return rates of AI toys, and the figures they arrived at cluster between 30% and 50%, all pointing to one fact: putting AI into toys is not hard; winning over users is. With opportunities and risks coexisting, at the 24th China Toy Expo currently being held in Shanghai, three exhibiting companies have made different choices. The China Toy Expo Trendy Toy Show, China Licensing Expo, China Baby Products Expo, and China Preschool Education Expo are being held concurrently in Shanghai, attracting 2,667 exhibitors, 5,504 brands, and more than 2,700 globally licensed IPs from 42 countries and regions. Renting a "brain" to make IPs From Hangzhou, Mituan Shuzhi showcased the not-yet-launched "Mizai · Five Mi Generals"—five AI cat-shaped companion robots modeled on Chinese rural cats, each corresponding to the coat pattern and personality of a real-life rural cat breed. For example, the tabby cat "Mi Lili" has a domineering-CEO persona: once it locks onto its owner, it only loves talking to its owner and tolerates everything about him; if anyone else tries to strike up a conversation with it, it will hiss and snap at them. The choice behind Mituan Shuzhi's product represents the attitude most toy companies currently hold toward AI: embrace it actively, but strictly control costs—if a third-party general-purpose large model can be called upon, never invest in R&D themselves. General Manager Shen Qiao'er admitted frankly that there are plenty of models on the market, and if the product design is not complicated, basically any of them can be used, so the company will not bear the training costs of foundation models and will instead concentrate resources on the "surface-level design" of trendy toy craftsmanship, character persona setting, and emotional interaction. The economic logic of this choice is very clear. When ByteDance released the Doubao large model in May 2024, it set the inference input price of the Doubao general model pro-32k version at 0.0008 yuan per thousand tokens, with the company stating officially that 0.8 li could process more than 1,500 Chinese characters. This signal at the time significantly lowered expectations for enterprise-level large-model calling costs. Companies do not need to sink costs into computing power and training, and the cycle from product definition to launch can be greatly compressed. But the price is equally obvious: when the underlying model is a public good, what products can compete on is the gap in shells, IPs, and marketing. If you are not a leading company, it is very hard to widen these gaps, and homogenization easily emerges, with a large number of products being "model wrappers." Mizai has achieved a certain degree of differentiation with its original rural-cat IP and emotional nurturing mechanism, but how long this product feature—stemming from design ingenuity rather than technological innovation—can stay "fresh" remains to be tested by the market. Five AI cat-shaped companion robots modeled on Chinese rural cats. Photo by Chen Xihan. "Five Mi Generals." Photo by Chen Xihan. Half self-developed "study companions" By comparison, Shanghai Miaoyin Habu has taken a bigger step, drawing a middle ground between "calling" and "self-developing." Their newly launched "Xiao Buren" and "Luan Dameng" this year are two "AI study companions," focused on the function of "accompanying children in reading." These products adopt a hybrid architecture of "Doubao + self-developed education model," with Doubao serving as the general dialogue and interaction foundation, while the self-developed education model solves what general models cannot. "During the product research stage, we found that most similar products are limited by copyright and cannot 'read' the textbooks children are currently using, so parents feel the product is muddling through and that their money is wasted." Ling Qiaodan, operations lead at Miaoyin Habu, revealed that on the basis of obtaining copyrights, the company has invested in its own R&D, with the self-developed portion accounting for 15% to 20% of total costs, mainly used for a self-developed AI children's agent, making the product sufficiently like a "study companion" that provides services such as textbook-synchronized read-along, spoken English practice, and mental math listening drills, with a sense of being a real person. The advantage of Miaoyin Habu's chosen path lies in deeply rooted scenarios: textbook copyrights, region-specific content adaptation, educational research review, and continuous updates are all advantages that cannot be directly obtained by renting ready-made large models, but the company must simultaneously manage models, content, hardware, and copyrights, resulting in higher initial investment and later maintenance costs. In addition, this choice carries other risks. Some industry practitioners point out that AI education toys easily fall into the trap of "pleasing neither side": on the one hand, being too educational may erode the companionship and soothing value of such toys; on the other hand, toys are always primarily for entertainment, and focusing on educational functions makes it harder to win the trust of the corresponding market. Ling Qiaodan revealed that the company has begun trying to "please both sides," and a bag-hanging product more oriented toward entertainment and companionship has already been developed. As for existing models, they will lean more toward education; once the specialized capabilities for educational scenarios are validated, the user-retention rationale they create will be harder for competitors to replicate than a plush shell. Two "AI study companions." Photo by Chen Xihan. Two "AI study companions." Photo by Chen Xihan. Building their own "brains" JD.com's path is not on the same dimension as the two startups: it does not make toys, but makes the toys' "brains." At the JD Supermarket booth, seven Kaimisi AI smart robot dogs danced the "floor dance" in response to voice commands. "The 'brains' of these robot dogs are all JD.com's self-developed solutions."Wang Hongkai, head of AI toys at JD Supermarket, introduced that JoyInside packages capabilities such as large models, speech recognition, dialogue management, long-term memory, and emotion perception into a ready-made solution for toy and hardware manufacturers. Manufacturers install this suite into their own products, and the toys can then chat, have memory, understand children's speech, and respond quickly. Currently, JoyInside has connected with more than 40 leading hardware brands, covering categories such as robots, robot dogs, and AI toys. A single brand does not need to build AI capabilities from scratch to obtain a dialogue experience close to that of leading players. Data shows that products connected to JoyInside have seen an average increase of more than 120% in dialogue turns, device response latency below 2 seconds, and children's speech recognition rates improved to over 95%. Wang Hongkai said that JoyInside provides a path to participate in competition without having to beat rivals on model capability. The model may not be the strongest, but if hardware manufacturers want to sell products, JD.com's combination of channels, supply chain, and free AI capabilities is a hard option to bypass. But this path is not perfect either. What JoyInside provides is general capability, not deep customization for a specific category, age group, or learning scenario. It is more like "pre-made AI dishes," which can make a plush toy able to chat and have memory, but cannot solve for the brand the question of whether "this toy understands its users." If the brand itself does not further refine scenarios and content, dishes cooked in the same JD.com "pot" will "taste" the same and lack their own vitality. In addition, over-reliance on voice interaction also easily makes AI toys "all look the same." A report released by the China Internet Network Information Center in September this year shows that 76% of generative AI users mainly use AI to answer questions. When users' expectations of AI are highly anchored to the "Q&A" function by general-purpose tools, if AI toys also follow the trend and bet everything on "Q&A," it will be difficult for them to form sufficient retention reasons in competition with mobile AI assistants. The three choices themselves are neither right nor wrong; the real dividing line lies in whether companies clearly know what they are betting on and whether they have equipped themselves with the corresponding capabilities for that bet. Directly calling foundation models while expecting to build barriers through "intelligence" is a mismatch; partial in-house development without sustained investment in content review and copyright updates will end halfway; building a foundation while ignoring the product completeness of integrators means the larger the scale, the faster the overdraft. JoyInside provides a path to participate in competition without having to beat rivals on model capability. Photo by Chen Xihan. Seven Kaimisi AI smart robot dogs danced the "floor dance" in response to voice commands. Photo by Chen Xihan.