Artificial Intelligence | AI communism?

Protest against Big Tech and waste of resources in AI development: Could local use of public domain models solve this problem? Photo: AFP/JOE KLAMAR Hardly anyone would have expected a positive contribution to the topic of data protection from Mark Zuckerberg. And yet the founder and boss of Meta has become the poster boy for one of the most remarkable developments in the field of artificial intelligence. In the summer of 2023, the social media group released its in-house language model Llama under a free license. Llama thus became digital public domain. Thanks to the group's “community license,” it has since been possible to download the Llama models in different sizes and thus operate a data-efficient, private chatbot. Today, Llama is no longer the only option for using AI in this way. There are more than a dozen language models under free licenses to choose from.

There are different ways to release these million-dollar models. With a few exceptions, all versions and sizes of Llama, the French AI model Mistral and the Chinese DeepSeek are freely usable, as is the language model Nemotron from the graphics chip manufacturer Nvidia and Qwen from the Chinese IT group Alibaba. Other AI developers are releasing slimmed down versions of their otherwise closed top models under their own names. This applies, for example, to Gemma from Google, Phi from Microsoft and GPT-OSS from OpenAI.

Commercial developers usually only publish the so-called weights: billions of numerical values ​​that, in a complicated mathematical interaction, enable computers to explain quantum physics, write poems and generate images. This practice is called “Open Weight”. There is also the more extensive “open source” approach, in which AI developers also document how and with what data they trained their models. This transparency can be found primarily in the models of non-commercial players - such as Soofi, which is backed by a consortium of German universities and research centers.

Regardless of whether it is open source or just open weight: If the weights are released under a license, autonomous use of AI is possible. You can download the models and operate them locally - unlimitedly, free of charge and, above all, without a data trail. You can chat with a model on your own computer or smartphone using utility programs. No data leaves your device.

The fact that companies sometimes give away products is actually old hat in the digital world. For example, there are leading companies behind important versions of the alternative Linux operating system family or the popular blog and web software Wordpress. Through free licenses - legally binding sets of rules - they turn their digital products into public goods that can be used without restrictions and even cloned and further developed. The companies finance themselves through so-called open source business models. The main or basic product is freely available. Some customers who want to use it professionally still pay for non-free additional versions, support, consulting or cloud services. The creation of digital common goods is not based on an orientation towards the common good, but rather an established business strategy.

Digital capitalism often creates oppressive power relations, but sometimes they also seem communist.

Analogously, one could speak of “open weight business models” when it comes to commercial free AI. If companies release the weights of their models, it is clear that private individuals in particular will use them without any money flowing to the AI ​​developers. However, companies and public administrations may opt for the larger and better top version - if, as with Google, OpenAI and Microsoft, only slimmed down versions are released. Or they pay the AI ​​developers to help with local setup. They rely on expertise to adapt a model to internal company purposes or use tailored, non-free special versions.

Opening up a model can also help build trust, spread the word, and even improve the product. A freely usable AI model quickly finds a large number of users. If things go well, an ecosystem is formed in which a community develops additional applications without payment and contributes suggestions for improvement.

Under certain circumstances, the economic calculations behind such business models will not always work out. And it is conceivable that at least some developers will change their strategies. In spring 2026, Meta announced a new model family: Muse Spark will be installed in Whatsapp, Facebook and Instagram instead of Llama in the future, as Meta was dissatisfied with the quality of the previous language model. It is still unclear whether Muse Spark will be under a free license to the same extent. And the Chinese government is considering restricting the practice of domestic AI developers releasing their models on a large scale.

Ultimately, however, it doesn't matter why companies make their models available to the public. A comfortable situation has emerged for society: Anyone who wants to use AI locally has many options. You can download language models and chat with them using a free, layman-friendly AI utility program – such as LM Studio, Ollama or GPT4All. Isn't this the perfect solution to one of the big problems with artificial intelligence - the immense amounts of data created by the use of cloud AI?

Nope. Using local AI tends to be a scarcity economy. You have to live with sometimes considerable restrictions. On the one hand, not all model families and versions are freely accessible. On the other hand, the question always arises as to which model size actually runs on your device. Companies, organizations and state administrations may have their own data centers that can handle even the largest free model versions - or they may rent data protection-compliant data centers from Europe. The Berlin administrative AI BärGPT, for example, is based on the French open-weight model Mistral.

The situation is different when private individuals or small collectives want to operate AI models on standard computers. They only run comparatively tiny model versions. Some of them often produce nonsense in terms of content and language. Others, however, work well enough even in compressed form, at least for certain purposes. What the small language models can often easily handle are simple questions or chats about the meaning of life. Local AI is also suitable for some very specific tasks, such as translating texts moderately well.

“In general, you can say: the larger a model is, the more capable it is of answering complex questions,” says Dr. Nicolas Flores-Mr from Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS. He researches AI, helped initiate the construction of the European model Soofi and is one of the project leaders. However, size is not the only deciding factor. Rather, it has been shown that even small models can handle some tasks adequately: “There are use cases where it is important that a model is one of the best in the world, for example in the area of ​​complex science or when it is important to have as much detailed knowledge as possible. But many tasks can also be solved well by small models. This applies, for example, to combining documents, generating standardized texts and querying simple information. In addition, says Flores-Herr, models can be used well if they are provided with additional information. The AI ​​researcher recommends “trying around with open models and being pleasantly surprised as to what makes sense for which fields of application.”

So AI under free licenses does not solve all problems. It does not change the dominance and lack of transparency of the well-capitalized players who dominate the AI ​​market. After all, these are the ones who provide the most important free models. But local AI offers a solution to at least one problem: data protection. Anyone who doesn't use the models from Mistral, Google or Microsoft in the cloud, but rather locally, has to live with restrictions - but you can be sure: no one is listening. The data remains on your own device. The fact that this is possible is due to the sometimes paradoxical logic of digital capitalism. This often leads to oppressive power relations, but sometimes also to conditions that seem communistic: companies develop products for a lot of money - and happily turn them into common goods for reasons of competitive strategy.

The »nd.Genossenschaft« belongs to the people who make it possible: our readers and authors. They are the ones who ensure left-wing journalism for everyone with their contribution: without profit maximization, media companies or tech billionaires.

→ report independently and critically → make topics visible that would otherwise go unnoticed → make voices heard that are often ignored → counter disinformation with facts → initiate and deepen left-wing debates

“Pay voluntarily” now and support the financing of our solidarity-based newspaper. So that nd remains.