Models
Open Source AI
Also known as: open-weight AI
AI that is released openly, so anyone can download, run, study and build on it, as opposed to closed models you can only use through a company's service.
Open-source AI means models that are released openly, so anyone can download, run and build on them, instead of only being able to use them through a company's app or paid service. This gives people more control, privacy (you can run it yourself) and the ability to customise. It has created a huge global community improving and sharing models, and it is one of the big dividing lines in AI: open versus closed.
Open-source AI refers to models released openly rather than kept private. In practice there is a spectrum: - Open-weight: the model's weights are downloadable (e.g. Llama, Mistral), you can run, fine-tune and self-host, but the training data/code may not be shared. - Fully open-source: weights plus training data, code and full documentation, allowing complete reproduction.
Why it matters
- Control and privacy: run models on your own hardware; keep data in-house.
- Innovation: a global community builds specialised and improved versions.
- Debate: openly releasing powerful weights spreads benefits and scrutiny, but cannot be undone and complicates misuse prevention, a live safety and policy discussion.
"Open-source AI" spans a spectrum from open-weight (downloadable weights, often under licences that may restrict use, training data/code withheld) to fully open-source (weights, data, code and recipes enabling reproduction). The distinction matters legally and practically: many prominent "open" models (Meta Llama, Mistral) are open-weight rather than fully open, and the Open Source Initiative published a formal Open Source AI Definition (2024) to clarify terms.
The open pole underpins self-hosting, privacy, customisation (fine-tuning/LoRA), research access and cost control, and drives a large derivative ecosystem, countering concentration in closed frontier labs. Central tensions: irreversibility of releases and misuse potential vs transparency and distributed innovation (the "open release" safety debate); licence restrictions that qualify "openness"; and sustainability/governance of open projects. Open and closed models increasingly coexist, with organisations mixing them by task, sensitivity and cost.