Miniaturizing AI models is done via quantization and distillation. DeepSeek uses a combination of these techniques to create models that are both cost-effective and high performant.