How can AI hardware get rid of the fate of being left idle to gather dust? Taking weather as the entry point, Qituantuan creates trendy active companion toys designed for young people.
When the weather changes, a trendy toy by your side will take the initiative to care for you like an old friend; in the supporting App, the joys, sorrows, anger and happiness of the day will be recorded as an exclusive "emotional weather". This is a brand-new experience created by DovoSense for its TUNTUN WORLD Qi Tuantuan product line.
As a technology hardware company that integrates original IP and self-developed AI Agent system, DovoSense is materializing weather-themed characters, breaking the boundary between virtual and reality, and redefining the emotional companionship of human-machine partners. The project was launched in early 2026, made its debut at the CTE China Toy Expo in Shanghai in October, and the product has now entered the Engineering Validation Test (EVT) phase with rapid iterations.
Using weather as the entry point to solve the retention challenge of AI hardware that "starts strong but fades quickly"
When founder Dong Siqi was conceiving the product, he focused on why users are still willing to communicate with AI after the novelty wears off. Qi Tuantuan targets young consumers aged 15 to 35 in first- and second-tier cities, and cuts into the scenarios of emotional companionship and personal expression. The team chose the weather theme because it changes every day, which naturally brings up topics, and it is easier for users to describe their indescribable moods through sunny, rainy, snowy and windy days. This design gives the character a natural reason to start a conversation, and allows users to record their moods in daily logs.
The team designed four original characters, Sunny Tuantuan, Snow Tuantuan, Thunder Tuantuan and Wind Tuantuan, around the themes of sunny, snowy, thundery and windy weather, and set unique personalities and expression styles for each of them. The same weather change can be talked about by characters of different personalities in their respective tones. The team is also developing a "Friend Guardian" function, which allows users to greet friends via weather-related messages, extending this kind of emotional expression between people.
The physical trendy toy and the standalone App undertake different interactive tasks. The hardware integrates AI voice and dynamic eye expressions, and the team is exploring the coordinated presentation of voice and visual feedback. The App is developed around weather services, mood records, friend guardianship and character interaction. The team hopes that the two entrances share the same character identity, memory and interaction status, so that users can continue the previous conversation on their mobile phones even if they do not bring the physical device with them when going out. The mood records in the mobile phone are also expected to become references for subsequent conversations. At present, the cross-terminal capability is still iterating with the prototypes, and the continuity during switching and long-term performance remain to be verified.
From "Passive Response" to "Getting Smarter the More You Use It": Building a Growing Companion with Layered Memory and Dynamic Persona
A greeting can start a conversation, and the next problems to solve are what the character remembers, how it responds, and when it is appropriate to speak again. The team is developing two sets of systems to handle the memories accumulated along with interactions and the long-term consistent character personality respectively.
Dovo-Mem is the layered memory system of Qi Tuantuan, which organizes recent conversations, important experiences and long-term preferences by time and usage. According to the team's design, recent conversations can help the character pick up topics that have just been discussed, and important experiences can be brought up again after multiple rounds of communication. The team is exploring how to retrieve and update memories according to the freshness of information and the relevance of topics. For example, if a user changes a preference, the system needs to use the new information the next time the relevant topic is discussed, so as to reduce repeated inquiries or the use of old impressions.
Dovo-Persona is responsible for the character's personality and expression. The team sets core personalities, speaking styles and expression boundaries for different characters, so that users can still recognize the familiar character from the way it speaks when they come back after a period of time. The character's response can be adjusted according to the situation and the user's communication habits. For example, the same character can follow up with more questions when the user wants to chat more, and stop in due time when the user prefers a short conversation. The team hopes that these changes are always consistent with the character's personality, and that the wording, voice and eye expressions convey consistent emotions.
On this basis, after the user's authorization, a continuously updatable user model will be formed by combining the actively recorded moods, historical communication preferences and interactive feedback, to distinguish the current emotion and the relatively stable communication preferences. Weather changes provide clues for active interaction, and the user's response to greetings and frequency settings can help the system gradually adjust when to contact, what to talk about, and how to express. The team hopes to judge whether a greeting is suitable for the current situation through subsequent feedback, and find a balance between active care and reducing disturbance.
Training a dedicated model for emotional companionship to build a long-term interactive data flywheel
According to the team, the project has completed the design of four original weather characters, developed the first batch of EVT prototypes, and built the technical architecture of the AI Agent system. The team is currently adjusting the hardware structure and interactive experience, coordinating trendy toy manufacturing, intelligent hardware and software development, and plans to push forward the Production Validation Test (PVT) before December 2026, and optimize the mass production cost at the same time.
In parallel with the R&D of the AI Agent system, the team is training a dedicated large model for emotional companionship scenarios, hoping to improve character expression, the use of historical memories and multi-round communication. This model has not been launched yet. Whether the character can accurately cite relevant memories after multiple rounds of communication and whether the speaking style can remain consistent remain to be evaluated and verified by subsequent products. The AI Agent system is responsible for organizing and retrieving memories, and the training of the dedicated model focuses on how to apply this information to responses that conform to the character's personality and fit the current situation.
The team hopes to filter the feedback from long-term interactions under the premise of user authorization and privacy protection, to evaluate and improve the product. For example, for the same active greeting, users may give different responses in different situations. Connecting the current situation, the character's response and subsequent feedback helps to compare different companionship methods, and may also accumulate materials for the training and evaluation of the dedicated model. For the team, the potential value of these materials lies in preserving the connection between shared experiences and subsequent reactions, and supplementing the interaction process that is difficult to present in a single round of conversation. Whether the data can support training and whether the improvements can be reflected in the product experience still needs to be verified through subsequent usage and evaluation.
Accumulate potential users through the App, drive hardware consumption through relationship guardianship
Different from the traditional path of selling trendy toys first and then building user connections, Qi Tuantuan plans to launch the App first, attract users with functions such as personalized weather services, mood records and friend guardianship, gradually build usage habits and character recognition, and then convert App users into intelligent hardware consumers.
Among them, "Friend Guardian" is an important entry point for the team's differentiated business model. Users can not only buy Qi Tuantuan for themselves, but also give it to friends or partners. Through the linkage between the physical trendy toy and the App, cross-space daily greetings and emotional care can be realized, extending hardware consumption to scenarios of gifting, two-person interaction and relationship guardianship.
In addition to the hardware, the team plans to launch doll costumes, limited editions and IP co-branded peripherals around the original weather characters to expand the derivative consumption of trendy toys. At the same time, as the App continues to iterate, the team will explore subscription services such as personalized interaction and mood review, and gradually form a business model that combines App customer acquisition, hardware conversion, IP derivative consumption and software value-added services.
The two founders have experience in marketing planning and system R&D respectively. Dong Siqi, Founder and CEO, holds a master's degree in international marketing from the UK, has been engaged in strategic planning at two listed companies, and participated in service projects for enterprises including Huawei and Tencent. Zhou Xuanchen, Co-founder and CTO, graduated from the University of Tsukuba in Japan with a major in computer science, has 10 years of experience in program development and system architecture, and 5 years of experience in AI development. He has published international academic papers in the AI field and participated in the review of AI journals, and is also an early contributor to open source projects such as OpenClaw and Hermes Agent.
In terms of channels, the team plans to accumulate early users through the App and domestic content platforms, promote hardware sales in combination with e-commerce platforms, trendy toy collection stores and technology retail stores, and explore cooperation opportunities in the Japanese, European and American markets. At present, the project is seeking financing, and simultaneously connecting with domestic and foreign channels and IP partners.