MakazhanAlpamys/Soup
Fine tunes a large language model from a single YAML file, with no SSH into a training box and no framework configuration to assemble. It streams layers rather than loading the whole model, which is what lets an 8 billion parameter model train on a laptop GPU with 4 GB of memory. Distributed on PyPI as soup-cli, so installs are countable separately from stars.
68 open issues against 430 stars is a 16% ratio, five times what a young repo usually carries, and forks are up 8.5% against 4.9% star growth. The people arriving are running it and hitting walls rather than bookmarking it.