The growing demand for urban air mobility, drone logistics, and low-altitude aerial services has highlighted the need for intelligent and autonomous aerial networking solutions. Low Altitude Intelligent Networking (LAIN) emerges as a promising paradigm. Large Language Models (LLMs), with their capabilities in reasoning, language understanding, and task planning, have the potential to significantly improve the autonomy of unmanned aerial vehicle (UAV). However, cloud-based LLM access is often impractical in LAIN due to connectivity limitations and latency requirements. Deploying LLMs at the edge node offers a compelling alternative by enabling real-time responsiveness and onboard intelligence, but poses challenges such as limited computation, energy, and memory. In this paper, we present an offloading optimization problem for LLM inference in LAIN, aiming to maximize inference throughput. The problem is a variant of multidimensional knapsack problem, which is NP-hard. To address this NP-hard problem, we develop Dynamic Batching with Genetic Algorithm (DBGA) to address the multidimensional knapsack problem with constraints include communication and memory resources on edge server and UAV-specific latency and accuracy demands. Simulation results indicate that DBGA surpasses other batching benchmarks in request completion radio across diverse settings.