Docs

The thesis, the bets, and how to rebuild the results.

Manifesto

openbox-llm decouples a model's effective capability from what must reside in GPU VRAM. Each bet spends an abundant resource — compute, host RAM, disk, an explicit index — to relieve the scarce one.

• Native sparse attention — validated. Compression, selection, and sliding-window branches cut attention's O(T²) cost.
• Trainable memory layers — capacity as addressable entries rather than dense parameters.
• Hot/cold offload — hot weights in VRAM, cold in host RAM or disk, swapped by explicit top-k selection.
• Knowledge injection vs. retrieval — inject document knowledge into memory, editable and inspectable; the open question is faithful recall at scale.

Full text: MANIFESTO.md

Get started

git clone https://github.com/msb-msb/openbox-llm
cd openbox-llm
pip install -r requirements.txt
python test_integration.py   # gate the kernels

Full instructions: README.md

Paper

The methods and results are written up in the Open-Box paper.