On-device Human Activity Recognition (HAR) requires balancing accuracy and deployment efficiency on constrained hardware. We present Lightweight Human Activity Recognition (L-HAR), a controlled comparison of Baseline and Lightweight Temporal Convolutional Network (TCN), Transformer, and Long Short-Term Memory (LSTM) models using data sampled at $100\mathrm{Hz}$ from three inertial measurement units (IMUs) worn by one participant. Evaluations covered classification, model footprint, multiply–accumulate operations (MACs), offline CPU latency, Desktop/Raspberry Pi streaming, 8-bit integer (INT8) quantization, and software-estimated Raspberry Pi energy efficiency. Lightweighting reduced parameters by up to $95.5\%$, model size by $94.0\%$, and MACs by $87.6$–$95.5\%$, with accuracy reductions of $0.8$–$2.2$ percentage points (pp) and sub-millisecond offline latency for all Lightweight models. Baseline and Lightweight models sustained approximately $100\mathrm{Hz}$ message/inference rates on Desktop, whereas no Raspberry Pi configuration reached $100\mathrm{Hz}$ inference. Among Lightweight models, Transformer achieved $9.381\mathrm{ms}$ End-to-End (E2E) latency and $66.925\mathrm{Hz}$ inference. INT8 quantization changed accuracy and F1-score by less than $0.13$ pp; Quantized TCN achieved the best Raspberry Pi streaming result ($9.066\mathrm{ms}$ E2E; $69.928\mathrm{Hz}$) with an estimated $49.933\mathrm{mJ}$ per inference. These results demonstrate architecture- and backend-dependent deployment behavior and the need for direct target-platform evaluation.