Open vs Closed AI Models

The difference between AI models whose inner workings are released publicly and those kept private, and the economic and safety arguments on each side.

AI developers make a key choice: release their models openly, or keep them private. This choice shapes competition, innovation and safety.

Closed models

Closed models are kept private by their developers. Users access them through apps or paid connections, called APIs, but cannot see or change the model’s inner workings. OpenAI’s GPT models and Google’s Gemini are examples.

Advantages for developers include controlling how models are used, earning revenue from access and protecting their investment.

Open-weight models

Open-weight models release the model’s weights, the learned numerical parameters, so anyone can download, run and modify them. Meta’s Llama models, and models from French company Mistral and Chinese companies such as DeepSeek and Alibaba’s Qwen, have been released this way, with varying licence terms.

Economic arguments for openness

Competition: open models let start-ups, researchers and smaller countries build AI without depending on a few large firms.
Lower costs: users can run models on their own computers or cheaper cloud services.
Innovation: many developers can improve and adapt models, like open-source software.
Sovereignty: countries and companies can keep data and AI in their own control.

Arguments for caution

Misuse: once released, open models cannot be recalled, and bad actors could remove safety features.
Revenue: releasing models may make it harder to recover huge training costs.
Safety concerns about very powerful future models.

A business strategy

Some companies release open models strategically, for example to build an ecosystem around their technology, attract developers, or reduce the market power of rivals whose business depends on closed models. This resembles the use of open-source software as a complementary products strategy.

An Indian language model

A research group wants an AI model that works well in Tamil. Rather than training a model from scratch, which would cost a fortune, it downloads an open-weight model and fine-tunes it on Tamil text at modest cost. Open models make this possible, lowering barriers for languages and communities that large companies may neglect.

Thinking open models are always free of restrictions

Many "open" models are released under licences with conditions, such as limits on use by very large companies or for certain purposes. Openness varies from model to model.

Key takeaways

Closed models are kept private and accessed through apps or APIs.
Open-weight models release their parameters for anyone to use and adapt.
Openness supports competition, lower costs and sovereignty.
Critics worry about misuse and recovering training costs.