Kristijonas Adomaitis has published Kadmos-1 on GitHub, a decoder-only transformer language model with 45.8 million parameters that holds a short conversation and writes small Python functions. The architecture, tokenizer and weights were built in the repository from random initialisation, not fine-tuned from an existing pretrained model. The model file is about 110 lines of PyTorch. The name comes from Kadmos, the figure in Greek myth credited with bringing the alphabet to Greece.

It runs anywhere PyTorch runs, using Metal (MPS) on Apple Silicon and the CPU otherwise. Download kadmos-1.pt from the latest release into the repository folder and start a chat. It keeps the most recent turns of the conversation within a 512-token context window.

Adomaitis is direct about the limits. At this size it often answers the wrong question, invents facts, and writes code that does not run. In the posted Python transcript, three functions run and two are correct, while the third returns a sum with wrong comments.…