The first model in the family and the first rung of a ladder that runs G-Micro, G-Mini, G-Core, G-Mega. Nothing here is a fine-tune of somebody else's weights: the tokenizer, the Polish corpus and the training loop were all built for it, which was the whole point, the aim was to have a model that works and to understand every layer of why it does.

It runs offline on a MacBook. Training happens on free cloud GPUs, because the hardware here has neither the memory nor the tensor cores to do it in any reasonable time.