Open Weights versus Closed Source AI: explanation and impact

AI models that can be freely downloaded are becoming more powerful, smaller and cheaper. Do companies still need an expensive data-center model for every AI task?

AI models that can be freely downloaded are rapidly becoming more powerful, smaller and cheaper. That raises a new question: do companies still need an expensive model running in a massive data center for every AI task?

To understand this development, it is important to distinguish between open weights, open source and closed source. These terms are often used interchangeably, but they do not mean the same thing.

What are open weights?

During the training of an AI model, billions or even trillions of numerical values are computed. These values are called weights. They determine how the model processes information and formulates an answer.

With an open-weights model, these trained values are made available for download. Developers and organizations can therefore run the model on their own computers or servers. They can also adapt it to a specific application, without sending every question to the original provider.

Open weights offer organizations, among other things:

• more control over the model;
• less dependence on a single vendor;
• the ability to use AI locally;
• more grip on sensitive data;
• possibilities to adapt the model to their own sector or task.

Open weights do not, however, automatically mean that a model is fully open source.

Open weights is not the same as open source

With open-source software, the source code is available to inspect, modify and redistribute. For a fully open AI system, more is needed than access to the weights alone.

The following components can also be relevant:

• the source code used to train the model;
• information about the training data used;
• the training method and settings;
• the evaluations and safety measures;
• a license that permits reuse and modification.

An AI model can therefore offer open weights while the training data and much of the development process remain closed. The license can also contain restrictions on commercial use or certain applications.

Open weights make a model usable and adaptable, but not necessarily fully transparent or reproducible.

What does closed source mean?

With a closed-source AI model, the weights, training data and technical approach largely remain in the hands of the developer. Users get access through a website, app or API.

The provider determines:

• which models are available;
• how much usage costs;
• which features are offered;
• how the model is secured;
• under what conditions data is processed;
• when a model is modified or removed.

Closed-source models offer great convenience. A user does not need to manage their own servers and typically gets automatic access to updates, support and built-in safety measures.

The flip side is that the user remains dependent on the provider. Prices, terms of use and access can change. It is also often harder to verify exactly how the model works and what happens to the information you enter.

Small models are often good enough

Open-weights models come in different sizes. Some can run on a laptop or even on a Raspberry Pi. Others are so large that they still need powerful, energy-intensive data centers.

Smaller models in particular can change the economic balance of the AI market. Many everyday AI tasks are relatively simple, such as:

• summarizing a document;
• classifying messages;
• extracting information from a text;
• drafting an email or report;
• searching internal documents;
• answering frequently asked questions.

For these tasks, the largest and most expensive model is not always necessary.

Stanford researchers showed with their Minions project how small local models can collaborate with a larger model in the cloud. In their tests, 97.9 percent of the quality of a full cloud solution was retained, while only 17.5 percent of the original cloud costs were needed.

According to research cited by FT Alphaville, small models performed as well as or better than larger models on 98.3 percent of the chat questions studied.

The exact results will differ per task. The main conclusion, however, is clear: many questions do not automatically need to be sent to the most powerful AI model available.

What does this mean for data centers?

In recent years, enormous sums have been invested in new data centers, chips and energy supplies. Those investments are based on the expectation that the use of large AI models will keep growing strongly.

Open-weights models can change that expectation.

When organizations perform a large share of their AI tasks locally or with smaller models, the need for expensive cloud processing per task declines. Large models and data centers remain necessary, but perhaps not for every summary, search query or first draft.

That does not mean data centers become redundant. Open models also need to be trained, and the largest variants require a lot of compute. What emerges is a more efficient division: small models for simple tasks and large models for problems where their extra capacity is truly needed.

The gap is narrowing at the top too

It is not only small models that are improving. Large models that may appear with open weights are also approaching the performance of the best closed-source systems.

Kimi K3 reached third place on the Artificial Analysis Intelligence Index in July 2026. That put the model on a par with some of the strongest closed-source models of that moment.

An important nuance applies here. The developer had announced its intention to release the weights of Kimi K3, but at the time of the analysis they were not yet available. Kimi K3 was therefore not yet an actually downloadable open-weights model.

The rapid development of models such as Kimi, DeepSeek and GLM does show how quickly a technological lead can disappear. A closed-source provider may deliver the best performance today, while a cheaper or more open alternative comes close a few months later.

Data from model platform OpenRouter points in the same direction. According to FT Alphaville, the share of questions for which a closed-source model offered the best combination of price and quality fell from about 60 percent at the start of the year to roughly a quarter.

Control over data and intellectual property

Cost is not the only reason to choose open weights. For law firms, research institutions, governments and companies with valuable intellectual property, control over data can be at least as important.

When an organization runs a model locally, confidential documents do not have to be sent to an external provider by default. The organization can decide for itself:

• where information is stored;
• who has access;
• which security rules apply;
• whether data may be used for other purposes;
• how long information is retained.

That makes open weights attractive for sectors working with trade secrets, personal data, legally privileged information or unpublished research.

A local model is not automatically safe, however. The organization remains responsible for security, updates, access control and testing the output. More control usually also means more technical responsibility.

The advantages of closed source remain

Open weights will probably not fully replace closed-source models. Closed providers can offer important advantages:

• easy adoption;
• professional support;
• regular updates;
• built-in safety measures;
• access to the newest features;
• strong performance on complex reasoning tasks;
• integration with other business services.

For many small organizations, a ready-to-use service is more practical than installing and managing a model themselves. Running an open-weights model requires technical knowledge, suitable hardware and ongoing maintenance.

The choice is therefore not only about performance. Organizations must also consider total costs, privacy, support, speed, reliability and the degree of control they need.

An AI market with multiple layers

The most likely outcome is not a market with one winner, but a market with several layers:

• small local models for everyday work;
• large open-weights models for organizations that want more control;
• closed-source frontier models for the hardest tasks;
• hybrid systems that automatically select the right model.

The real challenge for large AI providers is not that open weights make them immediately obsolete. It is that fewer and fewer questions require the largest and most expensive model.

When users can choose the smallest suitable model for every task, the pricing of closed-source providers comes under pressure. That also makes the economic case for unlimited data-center expansion less self-evident.

Open models complicate oversight

The spread of open weights also has consequences for the debate on AI safety.

In 2023, researchers and technology experts called in an open letter from the Future of Life Institute for a temporary pause in the development of systems more powerful than GPT-4. The goal was to create time for better safety protocols and independent oversight.

Open weights make that oversight more complicated. Advanced models no longer reside exclusively within a small number of large laboratories. Once the weights are published, copies can be stored, modified and redistributed worldwide.

Openness can promote research, innovation and independent scrutiny. At the same time, it becomes harder to centrally slow down the development or application of powerful models.

Conclusion

Open weights are changing the AI market because they make powerful technology available outside the infrastructure of large providers. They give organizations more control, can lower costs and enable local processing of sensitive information.

Closed-source models remain attractive because of their ease of use, support and high performance. But their position is less self-evident when smaller and more open alternatives can handle the majority of everyday tasks.

The future of AI will therefore probably not be fully open or fully closed. It will be a combination of local models, open-weights systems and closed-source services, where users decide per task which solution offers the best balance of quality, cost, privacy and control.