Published weights
An open source model is one whose parameters — its weights — are published, so anyone can download it and use or modify it. What is actually released and under what license varies quite a lot, which is why many people prefer the more precise term open weights.
How this differs from closed models
The models inside ChatGPT, Claude, and Gemini are closed, reachable only through the provider’s servers. An open model can be downloaded, which enables things a closed model cannot:
- Running it on your own machine or server as a local LLM, without sending data outside
- Fine-tuning it on your own data to build something specific to your work
- Using it freely for research, with an active community publishing improved versions
The main examples
Meta’s Llama series, models from Mistral in France, a group of strong models from Chinese labs, and Stable Diffusion for images are the best known. Hugging Face has become the de facto place where these models are shared, hosting an enormous number of them.
The performance gap against closed models has narrowed year over year, settling into a pattern where closed models push the frontier and open models broaden practical adoption.
Why it matters
Open models reduce dependence on a single vendor, lower cost, and make it possible to use AI without sending confidential data outside the organization. They also let researchers inspect what is inside, which matters for transparency and safety. Open versus closed is a recurring axis in industry news, and knowing the shape of the argument makes coverage much easier to follow.
Note that some models described as open source carry licenses with conditions on commercial use. For business use, check the terms of the specific model.