Here you go: DeepSeek R1 is an AI model. An AI model is a bunch of a matrices with floating point numbers (referred to as weights) where you feed in an input (a sequence of characters embedded as a vector of floating point numbers) and get an output sequence. DeepSeek is a mobile app (same name as the company) that lets you interact with that AI model through a chat interface. When you use their app, your data (prompts) go to their servers. The company has also open sourced (basically uploaded all those matrices) the weights of the AI model for free use by anyone. When you download those weights and bring it up yourself on your own server, you get to control the inference of the AI model and that way any user request sent to this new server doesn’t go to China as long as the servers are hosted in US. The weights are just a bunch of numbers organized as matrices executed with sequential matrix multiplies - so no computation needs to leave the server in order to compute the next word in a sequence. That way, another company can download the weights, host it on their servers, and let users interact with them in a chat frontend, and customize the AI model further to do more things like searching the web or using tools like code execution, wolfram, etc