thefinalboss/kahnn-nano

Kahnn nano (~1.8M) — CPU Chinchilla run + lifelong teach

Neuroscience-inspired non-Transformer model from AFKmoney/kahnn.

Files

File	What
ckpt_final.pt	Nano after ~37M tokens CPU pretrain (2026-09-09)
ckpt_teach.pt	Same + taught fact « La capitale du Canada est Ottawa. »

Code

All training / teach / forget code lives in the GitHub repo (not duplicated here):

• train_universal.py, teach.py, generate.py
• Docs: docs/UNIVERSAL_TRAINING.md, docs/LIFELONG_LEARNING.md, docs/CPU_RUN_LOG.md

git clone https://github.com/AFKmoney/kahnn
cd kahnn
pip install -r requirements.txt
# download weights from this Hub repo into ./runs/
python generate.py --checkpoint ./runs/nano_base/ckpt_final.pt --prompt "Once upon a time" --device cpu
python teach.py probe --text "capitale du Canada" --resume ./runs/teach/ckpt_teach.pt --config nano

Honest metrics (CPU box, 2026-09-09)

• ~36.99M tokens, ~1362 tok/s average at end, CE loss still ~10.5 (generation still weak)
• Teach exact fact → memory coherence ~1.0; partial probe ~0.43
• Prefer lifelong engram memory over fluent generation at this size/budget

License

See GitHub repo.