Artificial Intelligence launched on a microcontroller costing just 10 dollars

At a time when artificial intelligence technologies are developing rapidly, many rely only on expensive devices with powerful graphics accelerators and gigabytes of RAM. However, according to ixbt.com, Ukrainian developer Slava S (slvDev) broke all molds by successfully running a local language model on an ultra-cheap microcontroller costing only about 10 dollars. This is reported by Ixbt.com reports .

Within the framework of this project, named ESP32-AI, the entire process was carried out directly on the ESP32-S3 microcontroller without connecting to the internet or using cloud services. This achievement showed that in the future, artificial intelligence capabilities can be introduced even into the simplest household appliances and embedded systems.

A technical solution under resource constraints

At first glance, this task seemed almost impossible. The fact is that the ESP32-S3 is equipped with only 512 KB of internal RAM, 8 MB of PSRAM, and 16 MB of flash memory. Usually, such limited resources are considered insufficient even for the most compact language models.

To bypass this severe limitation, the developer used the Per-Layer Embeddings architectural approach applied in models of the Google Gemma family. After passing the model through a quantization process down to 4 bits, its size was successfully reduced to 14.9 MB.

The future and prospects of local intelligence

The largest table, containing about 25 million parameters, was placed in the flash memory, while SRAM was used exclusively for calculations. As a result, the demands on RAM were drastically reduced and stable system operation was ensured.

The model used in the project contains 28.9 million parameters and was trained on the TinyStories dataset. It is only capable of generating short fiction texts and is not designed for communication, coding, or searching for facts. Therefore, rather than replacing large language models entirely, the project clearly demonstrates the capabilities of modern architecture.

According to the author, this approach is very useful for autonomous devices where local data processing is important. In particular, home appliances, industrial controllers, and various embedded systems will be able to use artificial intelligence without needing a constant internet connection.