mmm1712/local_ai_esp32 is a repository for running a language model trained from scratch offline on the ESP32-S3. The latest release is v0.1.0-model-assets. The provided FACTS contain no specific details about what changed in this release, so the changes cannot be confirmed. The README calls the model PocketGPT. On XIAO you can talk to it with the touchscreen keyboard, and on DevKit over USB.
The focus of this project is running a small language model on a microcontroller without connecting to an internet service. Because it uses the provided weights and tokenizer, you do not need to train it yourself or prepare an API key. The repository is implemented in C and is licensed under MIT. It offers two input methods, screen and USB, but that difference does not change the model's capabilities. The project shows not only how to run the model, but also how to converse with the device.
According to the README, the model responds to short greetings and basic conversations. However, its answers may be wrong, repetitive, or unrelated to the question. Offline operation is therefore a verifiable feature, but it does not guarantee the accuracy or consistency of responses. There is also separate firmware that uses Gemini or Ollama, but that path does not use the PocketGPT weights. It is important not to confuse this with a response example of the PocketGPT model.