Cardputer AI offline chatbot

What you’re building

REZOR’s Cardputer AI runs the project’s quantized “8M” TinyTalk 2 model entirely on the ESP32-S3 inside the original M5Stack Cardputer and Cardputer ADV. Release 2.1 includes an app-only binary for M5Launcher and a full-flash image for esptool; both contain the model, tokenizer and interface, so the chatbot needs no Wi-Fi, API account or microSD card after flashing. The firmware selects the original Cardputer keyboard driver at boot, while its chat, story and raw-completion modes use the built-in keyboard and screen. The author reports about five tokens per second for the 8M model. Treat it as a constrained embedded demonstration, not a general assistant: context is about three short exchanges, useful replies top out around 250 tokens, factual knowledge is intentionally limited to simple trained examples, and output can still become nonsense. The release replaces the active firmware. Back up anything needed first, select the app-only or full-flash image for the chosen install method, and keep the project’s non-commercial dataset terms in mind before redistributing the bundled model.

Key steps

1. Back up any firmware or local data that must be kept, then confirm that the target is an original M5Stack Cardputer; release 2.1 also supports Cardputer ADV, but Flashable’s relationship here is for the original model.
2. Choose cardputer_ai_2.1.bin only for M5Launcher, which writes the app image at offset 0x10000, or choose cardputer_ai_2.1_full.bin for a complete esptool flash at offset 0x0. Do not interchange the two files.
3. For M5Launcher, copy or fetch the app-only binary and install it as replacement firmware. For esptool, put the Cardputer in download mode and write the full-flash binary at offset 0x0 over a data-capable USB connection.
4. Reboot and confirm that the firmware recognizes the keyboard and opens the chat interface without requesting Wi-Fi, an API key or model files on microSD.
5. Try a short chat, story prompt and /new reset. Judge the result within the documented limits: roughly three short exchanges of context, narrow factual knowledge and occasional nonsensical output.