Developer Alexey Borsky released Mini-AGI, an open source, MIT-licensed language model trained from scratch on an 8GB laptop GPU using a single-pass, batch-1 stream of data, according to the project's GitHub repository. The project reached the front page of Hacker News as a Show HN post.
The model reads text at the byte level and processes it through two dense prelude blocks followed by up to 24 applications of a recurrent block, each drawing its own eight experts from a shared pool, the README says. Expert weights sit on disk and page into VRAM as needed, with only 32 experts resident in memory at once, letting the total parameter pool, currently 540 million parameters, exceed what the 8GB card can hold at any moment.
To fight catastrophic forgetting, the project trains its core components, embeddings, attention and routers, at one-tenth the learning rate of individual experts, which the README says retains 99.84% of prior progress when the model learns a new subject compared with a model trained on that subject from chance. There is no frozen base model: training and inference run through identical code paths, and the model has processed 318.1 million characters across eight subjects so far.
Continual learning without forgetting is usually a problem attacked with large compute budgets and checkpoint comparisons. Doing it on a single consumer GPU, with the training loop itself as the only copy of the model that exists, is a different kind of demonstration: that the technique doesn't require infrastructure most individual developers don't have.