Natural language processing systems leveraging on-premises large language models (LLMs) can translate natural language into structured JSON commands for Space Traffic Management (STM) systems. While cloud-based LLMs excel at this task, security constraints necessitate local deployment, requiring evaluation of smaller on-premises models. We demonstrate that resource-efficient 7B-parameter models can achieve high accuracy for STM command generation through a two-stage pipeline. Our pipeline first classifies objectives, then generates schemas. Empirically, we observe that initial classification accuracy strongly influences overall performance, with failures cascading to the generation stage. We demonstrate that quantization disproportionately increases structural errors compared to semantic errors across 405 objectives. The best quantized model (Falcon3-7B-GPTQ) shows a 3.45% accuracy drop, primarily from structural errors. Our findings highlight limitations in how model compression affects applications that require syntactic validity. More broadly, we explore the feasibility of LLM deployment in air-gapped environments while uncovering how quantization asymmetrically impacts structured output generation.