MediLITE-QA: Energy-Efficient Medical Question Answering with a Quantized Domain-Specific Language Model on a Commodity Laptop CPU

Frontier large language models (LLMs) depend on energy-intensive cloud data centers and continuous connectivity that many low-resource healthcare settings lack. This paper investigates whether Low-Rank Adaptation (LoRA) fine-tuning combined with 4-bit GGUF quantization enables accurate medical question answering (QA) on commodity CPUs without GPU acceleration or cloud connectivity. We instantiate this pipeline as MediLITE-QA, a fine-tuned Phi-3.5-mini-instruct (3.8B-parameter) model. On a consumer-laptop CPU, the deployed model reaches a steady-state Time to First Token of 1.34 s, 8.19 tokens/second, and 71.87% average accuracy across nine medical QA benchmarks (72.79% before quantization, not significantly different from the 8B model OpenBioLLM-8B). The 7–8B comparison models used 1.56–1.98× more CPU energy per token and showed cold-start penalties that MediLITE-QA did not. Fine-tuning GPU-board energy, estimated from power telemetry, was approximately 139 kWh. In an exploratory blinded clinical evaluation, the fine-tuned model was not significantly better than its base model and remained significantly below a frontier cloud reference (Wilcoxon padj=0.0023, n=100). On this hardware tier, the pipeline is a viable, lower-energy-per-token alternative to larger models, relevant to Sustainable Development Goals 3, 9, and 10; results should not be assumed to transfer to phones, single-board computers, or servers, and accuracy gaps must close before clinical use.