AI glossary for businesses
Open-source AI models
In short
Open-source AI models are AI models whose weights, and sometimes code and training details, are published so anyone can download, run and adapt them. Examples include Meta's Llama, Mistral's open models, Google's Gemma and Alibaba's Qwen. Many are more accurately called "open-weight", because their licences or training data are not fully open.
How open-source AI models work
With a service like ChatGPT, the model runs on the vendor's servers and you send your text to it. With an open model, you download the model files and run them yourself, on your own computer, your own server or a cloud account you control.
An analogy: renting a serviced office versus furnishing your own. The serviced office is ready on day one and someone else handles maintenance. Your own space gives you control and privacy, but keeping it working is your job.
A note on terms. "Open source" strictly means the licence lets anyone use, study, change and share the work freely. Many popular models only publish the weights (the trained parameters) and come with licences that limit some uses. That is why experts often say "open-weight". Always read the licence before commercial use.
In practice for a small business
Most SMEs do not need to run their own models. Business plans of ChatGPT, Claude, Google Gemini or Microsoft 365 Copilot are simpler and usually perform better on general tasks. Open models start to make sense in a few situations:
a law or accounting firm wants to summarise sensitive documents without sending them to an outside service;
a manufacturer needs a narrow task, such as classifying technical notes, on a local machine with no internet connection;
a software partner builds a product and wants to adapt a model through fine-tuning.
Tools like Ollama let a technical person run smaller open models on a laptop for testing. Production use needs proper hardware and security. The article on whether to build, buy or customise AI helps frame the choice.
Why it matters and what to watch out for
Control is not automatic compliance. Keeping data in-house can help with GDPR, but you still need security, access control and a legal basis.
Hidden costs. Hardware, setup, updates and skilled people often cost a small team more than a subscription.
Quality varies. Smaller open models can be weaker in Greek. Test them with your own documents.
Licences differ. Some allow broad commercial use, others add conditions. Check each one.