Most commercial software is a sealed box. Open source is the opposite: the code is published for anyone to read, run and modify. It is not a fringe practice; much of the internet runs on open-source software, and companies large and small both use it and publish it.
In AI, the term has a special flavour. An "open" model is one whose weights (the trained brain itself) are released so anyone can run it on their own machines, rather than renting access through an API. Meta's Llama models and DeepSeek's releases are the famous examples, sitting opposite the closed models from OpenAI and Anthropic. Tools can be open source too: OpenClaw is open-source software wrapped around whichever model you connect.
Why you keep hearing it
Because open versus closed is one of the AI era's big fault lines: cost, control, safety and geopolitics all run through it. Every DeepSeek headline is really a story about open models catching up with closed ones.
What it means for you
Practically: "open source" on a tool means free to use and inspect, but also usually self-service, with more setup and no support line. For most marketing teams, paid closed products remain the comfortable default. Knowing the distinction mostly helps you follow the news.