Cheapest way to deploy smaller fine-tuned AI models?

Any tips on how to deploy and use a fine-tuned model on Huggingface in a cost effective way? Right now looking into use Gradio with HuggingFace spaces and using the API endpoint from there. Inference endpoints and Sagemaker seem excessive for this. The whole idea to use smaller models is to decrease costs (vs using a bigger model with an API endpoint) but maybe this just isn't cost effective for where we are right now.