The tech
An AI model learns from huge amounts of data and stores what it learned as numbers called weights. With an open weight model, those weights are free to download. You can run the model on your own computers and, if the licence allows, adapt it to your needs.
Closed models keep their weights private. You can only use them through the provider's app or paid service.
Open weight isn't the same as fully open source. A fully open source model also shares the code and details of the data it was trained on.
In technical terms: open weight releases the trained parameters for self-hosting and fine-tuning, while closed models are only served through a provider's API.
Why it matters
- Cost. Open weight models are often much cheaper per use than closed frontier models, and there are no subscription fees. That could make AI affordable for small businesses, local startups and public institutions. Running your own model does bring hardware, maintenance and staff costs, so compare the total.
- Modest hardware. Smaller open models can run on ordinary computers. A technique called quantisation reduces the memory a model needs by storing its numbers at lower precision, at some cost to quality.
- Control over data. Because the model can run on your own or regional servers, sensitive information doesn't have to be sent to a foreign tech company, and the system can keep working without a constant internet connection.
- Adapting to local needs. Developers can fine-tune an open model for regional, Indigenous and under-served languages, or for specialist terminology, that commercial services handle poorly.
In the real world
- Services in local languages. Jacaranda Health adapted Meta's open weight Llama models to build UlizaLlama, a Swahili-focused maternal health model. It powers PROMPTS, a free SMS service that answers pregnant women's and new mothers' questions, with nurses stepping in when a risk is flagged. PROMPTS has reached more than 2.7 million mothers and newborns in Kenya.
- Tools with limited connectivity. A suitable model running locally could help staff search programme guidance where internet access is unreliable.
- Locally managed systems. Public institutions, universities and local tech firms can build and maintain applications on infrastructure they control. Open weights aren't only for language: Farmer Lifeline Technologies in Kenya built a low-cost, solar-powered pest detection device on YOLOv5, an open computer vision model. It monitors fields, identifies pests and diseases, and sends alerts and advice to farmers by SMS. The company reports it can cut crop losses by up to 30% and raise yields by up to 40%.
The practical question is whether greater control justifies the running costs. A pilot should compare accuracy, total cost and maintenance against an existing service or way of working.