Towards Edge-deployable Telecom Intelligence with Efficient Small Language Models

Deployment of large language models in 6G and O-RAN edge environments is constrained by latency, hardware, and privacy requirements that make cloud-hosted inference impractical. While sub-2B small language models (SLMs) are suitable for local deployment, their general-purpose pretraining provides limited telecommunications-domain knowledge, and existing fine-tuning approaches overlook the diversity of knowledge areas and reasoning types the domain encompasses. We propose lightweight task-specific LoRA fine-tuning, training one compact adapter per task category over a shared frozen backbone, and evaluate on the GSMA ot-full benchmark across six task categories using three SLMs. Our results show that SLMs with lightweight LoRA adapters not only narrow the performance gap with larger models but occasionally surpass them, while each adapter introduces only a small fraction of additional parameters relative to the frozen backbone. These findings suggest that lightweight task-specific adaptation offers a practical and efficient path toward edge-native telecommunications intelligence.