On-device (offline) quantizations of thesreedath/slm-125m-raft-dpo, our best 125M model with RAG (standard Llama architecture, 12 layers, 768 hidden, 16K vocab). Load with any llama.cpp runtime (Ollama, PocketPal AI, LM Studio, ChatterUI).

For RAG, prepend retrieved passages as context blocks before the question.