Local models are language models run on a company's own hardware instead of a cloud service. The main reason for deploying them is that the processed data never leaves the company's infrastructure, which addresses concerns around sensitive documents and in regulated industries. A secondary reason tends to be predictability of costs at high processing volumes. The price for this is real, however: the necessary hardware is not cheap, open models tend to be weaker than the best commercial ones on more complex tasks, and the company takes on responsibility for operation, updates and security. In practice, local models pay off for high-volume, lower-complexity tasks – classification, data extraction, transcription – and less so for tasks requiring the highest quality of output.
See also: Cloud AI platforms, How much does an AI solution cost, RAG over a company's own content.
