TinyTalk 2
A chat model small enough for a microcontroller. It makes small talk, tells short stories and knows when to stop talking.
Most chat models need a graphics card and gigabytes of memory, so a chatbot usually has to call a server. TinyTalk 2 has 8 million parameters. It is small enough to run on a microcontroller, but it runs on anything larger, too.
It can do small talk, remembers the last few messages, and writes short children's stories or answers simple questions about colours, animal sounds, and opposites. If a question is too complex, it says "I don't know" instead of making things up. Small models tend to ramble or repeat what you say, so I trained this one to answer and then stop.
Use it as the brain for your own device, a way to test an AI pipeline without a massive download, or a tool to learn how tiny models work. It runs in Ollama and llama.cpp.