Open weight vs open source: why precision matters for your SME

In the world of artificial intelligence, the term open source is used with a lightness that worries experts. Peter Farkas, CEO of Percona, has made a clear request from the Open Source Summit Europe in Prague: do not use open weight and open source as synonyms. The distinction is not trivial, and even less so for an SME that wants to build on open technology without surprises.

What an 'open weight' model really is

When a model is published as open weight, what is released are the weights: the numerical parameters that the model has learned during its training. That allows you to download it and run it on your own infrastructure. But, as Farkas points out, you do not have the source code or the training data; only the result of both. Therefore, you cannot reproduce the model from scratch or fully audit how it has been built.

Confusion is frequent. Models like those from DeepSeek are often described as open source, when in reality they only publish the weights. Farkas insists that this is not bad: open weights are valuable because they allow experimentation and local deployment. The problem arises when they are presented as equivalent to traditional open source, because then the meaning of the freedoms it guarantees is diluted.

The difference that matters: running vs. trusting and improving

James Landay, director of the HAI Institute at Stanford, summarizes it with a question: open weights answer “can I run this?”, while open source answers “can I trust this, improve it, and build the next thing on it?”. For an SME, the difference is practical. If you only have the weights, you can use the model, but you cannot deeply adapt it or verify its internal functioning. If you also have the code and the data, you can audit, modify, and share improvements, which reduces dependence on the provider.

Farkas warns that, without a clear definition, open washing wins: companies that label as open something that is not. And that can affect the definition of open source in software in general. Imagine that a tool is advertised as Apache 2.0 but has usage restrictions in the European Union. That kind of ambiguity is what he wants to avoid.

The debate on the definition of open source AI

The Open Source Initiative (OSI) published in 2024 its first definition of open source AI, with criteria on the freedoms to use, study, modify, and share AI systems. But the definition is still under debate, especially regarding what information about training data should be revealed. Duane O'Brien, new executive director of OSI, acknowledged the criticisms and announced that they are reopening the conversation. They have launched a two-year fellowship and plan a series of community discussions to review the definition.

Meanwhile, some manufacturers go further. Xiaomi, for example, has published not only the weights of its MiMo-V2.6 models, but also reinforcement training environments, code, and documentation. This shows that there is a spectrum: not all “open” releases reveal the same thing.

What this means for your SME

If you are evaluating AI models to integrate into your business, do not be carried away by the open source label. Ask yourself what you really need:

• Do you just want to run the model on your servers? An open weight model may be enough. You will be able to experiment and use it in production if you assume the risks.
• Do you need to audit it, modify it, or integrate it deeply? Then look for true open source: code, training data, and licenses that allow those freedoms.
• Are you concerned about dependence? A model with only weights ties you to the provider for improvements or fixes. With open source, your team can intervene.

At ForgeNEX we recommend always reading the fine print of the license and checking what is exactly released. A model that you can use in your SME is not the same as one that you can adapt and share without restrictions. Clarity in the terms will save you legal and technical problems.

Furthermore, this debate connects with other trends that we have already analyzed, such as the end of fidelity to a single AI model. SMEs are diversifying providers, and understanding what each license implies is key to not getting trapped.

Conclusion: demand clarity

Farkas's request is not a semantic whim. Behind the words there are concrete freedoms that affect your ability to innovate and control your technology. As an SME, you have the right to know if a model is truly open source or only open weight. Ask, read the licenses, and choose accordingly. Precision in language is the first step to making solid technological decisions.

Source: The New Stack. Analysis and adaptation: ForgeNEX.