Latency is a crucial challenge when deploying language and AI models in production. Ignoring computational performance, focusing solely on accuracy, is a mistake. Speed, measured in tokens per second, has become an essential architectural metric. This is because critical applications require fast responses, especially in finance and healthcare. Optimizing model performance also leads to lower GPU utilization and, consequently, lower cloud bills. Techniques such as using lightweight architectures, quantization, and batch processing help improve speed without sacrificing accuracy. A practical example in Python is presented to measure model speed in tokens per second. A target of approximately one hundred tokens per second is a good benchmark for real-time human interactions. Steps for optimization include establishing an initial benchmark, implementing batch processing, and evaluating distilled or quantized models. The community shares tools and libraries for profiling and benchmarking AI models.