The first time many parents cut open a talking plush bear, they discover that the "smart" does not come from a Bluetooth speaker but from an AI toy module stuffed into a cavity that used to hold nothing but stuffing. At Shenzhen Nablai Intelligent Technology Co., Ltd. (Nablai), three product family codes - TY (based on Tuya T5E), LX (based on Espressif ESP32S3) and NT (based on PY32) - are how we turn "making a toy talk" into an embedded module that can actually be mass produced. So what really is inside this little thing?
Main control: the brain of the module At the heart of the module sits the main control MCU. The TY Series uses the Tuya T5E, the LX Series the Espressif ESP32S3, and the NT Series the PY32. A genuine AI toy module runs a dual stack of "local SDK + cloud API": the local SDK handles audio capture, the wake word and servo control, while the cloud API handles LLM inference and content updates. TY002 and LX002 both sit on this dual-stack architecture; the only difference between them is ecosystem ownership and brand sovereignty - TY draws on the full Tuya stack and works out of the box, while LX pairs the Espressif ESP32S3 with the Nablai private cloud to give the brand full autonomy. NanoToy (PY32) is the true pure-UART command module: no semantics, no cloud, no SDK, built for low-cost trial runs.
Hearing: dual-microphone array and audio codec For a toy to hear a child clearly across a room, the module needs a dual-microphone array plus an audio codec for noise reduction and beamforming. We are always reminding customers of one thing: the position of the microphone openings and the thickness of the plush have a direct effect on pickup, so this has to be settled together with the exterior design during prototyping - changing the cavity later is expensive.
Sound: speaker and acoustic cavity The same speaker, buried under different thicknesses of plush, can come out noticeably different in loudness and timbre. A module vendor has to prototype and calibrate the acoustic cavity over and over again - engineering know-how that nobody can copy in the short term. On the LX Series we tune the voice to the IP character, so that it is "the character speaking" rather than "a generic robot voice".
Connectivity: Bluetooth and Wi-Fi modules The module gets online through a Bluetooth/Wi-Fi module: voice goes up to the cloud for inference, and content comes back down. The LX Series ships with the Nablai app, WeChat Mini Program and the AMS vendor management platform, with accounts, subscriptions, content and data owned 100% by the brand owner; the TY Series rides on the Tuya ecosystem, with pre-compliance for overseas markets and 60+ languages out of the box.
Power: Li-ion battery and power management A toy has to be rechargeable and it has to be safe, so the module carries a Li-ion cell and a power management system. Size and battery life trade off against each other, and at the prototyping stage we propose a battery solution matched to the usage scenario - long companionship sessions versus short bursts of interaction.
Structure: fitting the silicon into a stuffing cavity Getting all of these parts into a cavity that used to hold nothing but stuffing places real demands on structure, heat dissipation and mounting. A module has never been "buy a chip and stick it on" - it is a systems engineering job. What we normally deliver is an all-in-one module with the interfaces closed out by a single vendor, which avoids the integration finger-pointing that comes from sourcing voice, servos and cloud from three different places.
Why a dual stack and not "API vs SDK, two options" Every so often the industry writes the architecture up as two parallel options, "one cloud API line and one local SDK line", and that is not accurate. Both TY002 and LX002 are a local SDK + cloud API dual stack; the difference lies in ecosystem ownership and brand sovereignty, not in "one being purely cloud and the other purely local". NanoToy is the true pure-UART command module: it runs scripted command performances for the expression screen, vibration and audio, and does no semantic understanding at all. Do not conflate the three when you make your selection.
Policy is pushing "embeddable", too This is turning into an industry trend. Under Guangdong province's Action Plan for Accelerating AI Adoption in the Toy Industry (2025-2027), the province is targeting RMB 100 billion in revenue across above-scale toy enterprises and AI toy penetration above 30% by 2027, and the plan explicitly supports AI chips, AI modules and toy-specific on-device models that are easy to embed. Separately, in an on-the-ground survey of the Guangzhou toy market carried out by Leikeji in early 2026, AI toys generate a great deal of online buzz but still have low offline distribution and weak consumer awareness, with shipments currently going mainly to overseas markets such as Europe and the Middle East. Turning AI into "a module that can be embedded in a toy" is precisely the key path to closing the gap between awareness and supply.
Engineering notes for toy factories
- Settle the acoustics first: confirm the microphone openings and the speaker cavity together with the exterior design during the prototyping window.
- Yield is a system property: the module is a systems engineering job, and yield of 99%+ comes from refining structure, acoustics and integration together as a whole.
- Close out the interfaces: module, app and cloud are one solution, and assembling them separately makes later integration more expensive.
- Align the schedule: standard modules are held in stock, 2-3 weeks of prototyping locks the BOM, and mass production then transitions smoothly to weekly delivery.
Module versus AI box: why not just use an external device Some customers ask why we do not simply hand the child a talking AI box. The reason is that companionship with a toy happens in the "held in your arms" scenario; an external box puts something between the child and the toy, and the interaction breaks. Making intelligence a module lets it live on the toy itself - the child talks to the plush, not to a bolt-on gadget. This is also why we insist on building embeddable modules rather than selling a standalone piece of hardware.
Certification and safety The module involves radio transmission and use by children, so certification cannot be skipped. The TY Series draws on the Tuya ecosystem and comes with CE/FCC and pre-compliance for overseas markets; the LX Series is configured for the certifications the target market requires. We fold the certification scope into the structural assessment during prototyping, to avoid rework once mass production has started.
FAQ Q1: How small can a module be? Most integrate into a standard doll cavity without breaking up the existing look. Q2: Can we keep our own brand? Yes. On the LX Series, 100% of the data belongs to the brand owner; through the Tuya ecosystem, the TY Series can also be white-labeled for overseas markets quickly. Q3: How is content updated? LX runs on our own AMS, which gives the IP owner complete control over scripts, voice and update cadence; TY is operated and maintained centrally by the Tuya platform. Q4: Roughly how long is the lead time? Prototyping usually takes 2-3 weeks, and a typical custom project reaches mass production in 3-4 months.
About Nablai
Nablai is an AI company, and AI smart toys are the application area we focus on today. Our product is the AI smart toy module, and our goal is to give toy companies a product-level solution for bringing AI into their toys.
We have built three product families:
- TY Series: built on Tuya T5E, 60+ languages, the fast lane into global markets
- LX Series: built on Espressif ESP32S3, full brand ownership, deep customization
- NT Series: built on PY32, a low-cost way to test the smart toy market
- Website: https://www.nablai.com.cn
- Phone: 15917703502
- Slogan: Driven by gradients, optimized by intelligence. Powered by Nabla, Optimized by AI.
Shenzhen Nablai Intelligent Technology Co., Ltd. — AI module specialists for smart toys
📧 contact@nablai.com.cn 🌐 www.nablai.com.cn