Volotat/mini-AGI-undertrained

mini-AGI — undertrained snapshot

These weights are undertrained. They are a snapshot of a training run that is still reading its first pass over the corpus. Once that pass is complete, the final weights will be published, and this repository may be deleted.

Snapshot: 12 Oct 2026 · 1,872M characters read (24% of the first pass) · 397 experts

mini-AGI is a small continual-learning, byte-level language model. It keeps its weights as ordinary files on disk and pages them onto the GPU as it needs them, so it trains and runs on a single 8 GB card, grows new experts when it runs short and prunes the ones nothing asks for. It is a toy-level experiment, not a frontier model.

Code, architecture and the full training history: volotat/mini-AGI

Use it

1. Get the code and install it as the GitHub README describes.
2. Download these weights into it (about 10 GB):

   cd mini-AGI
   hf download Volotat/mini-AGI-undertrained --include "weights/*" --local-dir .
   

3. Talk to it:

   python3 serve.py --port 8080     # then open http://127.0.0.1:8080
   

   It learns from your conversations and saves what it learned into weights/. Add --no-learn to keep the snapshot as it is.
4. Or keep training it on your own text:

   python3 train.py read path/to/your/text --save --weights-dir weights
   

What is in weights/

File	What it holds
manifest.json	what the model is: its shape, its experts, how much it has read
core.npz	the shared body of the network
routers.npz	how the model picks experts, and how loudly each one speaks
experts/	one file per expert, with its optimizer state
optim.npz	optimizer state for the shared body

The optimizer state is included so that training can continue exactly where this snapshot stopped.

License

MIT