Despite advances in large-scale language models (LLMs), low-resource languages remain underrepresented in NLP, limiting digital accessibility for millions. To address this, we present PunGPT2, a fully open-source generative model suite tailored for Punjabi. Trained on a 35GB corpus of literature, religious texts, news, and social discourse, it captures the syntactic and morphological richness of Punjabi through tokenizers optimized for Gurmukhi and Shahmukhi scripts. We introduce Pun-RAG, a retrieval augmentation framework that integrates PunGPT2 with the FAISS retriever, and Pun-Instruct, which uses QLoRA for instruction-tuned zero-shot summarization, translation, and question answering. Furthermore, we develop Quantum-RAG, which fuses sparse, dense, and quantum kernel embeddings to enable efficient, context-aware retrieval with low memory overhead, marking the first practical implementation of quantum-inspired retrieval in low-resource LLMs. This model outperforms multilingual baselines (mBERT, mT5, MuRIL, BLOOM) on FLORES-200, IndicGenBench, and the new PunjabiEval suite. Quantum-RAG achieves +7.4 Recall@10 over FAISS and +3.5 BLEU over mT5 on PunjabiEval. By releasing the 35GB Punjabi corpus, the PunjabiEval benchmark, all model weights, training scripts, hyperparameters, and evaluation pipeline, we establish a new state-of-the-art in Punjabi generation and retrieval.
