This version is somewhat experimental - it trains significantly less of the model while keeping the same rank, hence it's absolutely tiny file size. It should be less overwhelming than V1.