Open-weights model

What open weights means exactly

Technically, a large language model consists of two things: an architecture (the blueprint of the network, such as a transformer with a certain number of layers) and the weights, i.e. the billions of numerical values that were learned during training. The architecture is usually described in scientific papers and is easy to reproduce. The real value lies in the weights, because training them costs millions to hundreds of millions of US dollars in computing time, depending on the size of the model.

With an open-weights model, the manufacturer publishes precisely these weights, usually as files in Safetensors or GGUF format on platforms such as Hugging Face. Anyone who downloads them can start the model locally with inference software such as vLLM, llama.cpp or Ollama, without calling the provider's API. The data then does not leave your own data centre, and the operator pays no token prices, only for their hardware.

Three levels of openness

In practice, three levels have become established that are often lumped together:

The Open Source Initiative's Open Source AI Definition (OSAID), published as version 1.0 in October 2024, draws the line clearly: a model only counts as open-source AI if, in addition to the weights, the code and a sufficiently precise description of the training data are also open, so that a skilled person could rebuild the system. Most popular "open" models do not meet this condition. They are open weights, not open source.

Licences: the real difference

A licence is always published with the weights, and this is where the wheat is separated from the chaff. For business use, the licence is more important than the question of whether the training data is open.

• Free software licences (Apache 2.0, MIT): commercial use, modification and redistribution are permitted without restriction. Mistral, Qwen (in most sizes), DeepSeek and OpenAI's gpt-oss use such licences. For a company, this is the most straightforward variant.
• Community licences with conditions: Meta's Llama licence allows commercial use but requires a separate agreement above a threshold of 700 million monthly active users and mandates attribution. Such licences are not open source in the OSI sense, but are practically usable for the vast majority of companies.
• Licences tiered by revenue or usage: some providers link free use to revenue limits of the licensee, above which a commercial licence becomes due. Here the legal department must read carefully.
• Research-only licences: use for research only, no commercial use. Unsuitable for use in a company.

A common mistake in projects: a model is planned in as "open source", and only shortly before go-live does it become apparent that the licence contains a usage restriction or attaches conditions to the redistribution of derived models (for example after fine-tuning). Checking the licence therefore belongs at the beginning of every model selection, not at the end.

Why open-weights models are relevant for mid-sized companies

For companies in German-speaking countries, open-weights models are interesting for four reasons that have little to do with ideology and a lot to do with business economics and law.

Data sovereignty and the GDPR

Anyone who runs a model on their own infrastructure or with a European hosting provider does not send any data to a US or Asian provider. This considerably simplifies the data protection impact assessment, makes many data processing agreements superfluous and defuses the third-country issue. For industries with special confidentiality obligations (healthcare, legal advice, financial services, public administration), this is often the decisive reason.

Cost structure

Closed models are billed per token. With high, even volumes (such as classifying support tickets, enriching product data or extracting documents), a self-operated open-weights model can be considerably cheaper, because only hardware and operation costs arise. With low or heavily fluctuating volumes, the calculation is reversed: then the API is cheaper, because there is no need to pay for an idle GPU.

Control over versions

API providers regularly replace models and switch off old versions after a transition period. A process that has been tuned to a particular model version can change unnoticed as a result. Downloaded weights do not change. The company decides for itself when to migrate and tests beforehand.

Adaptability

Only with access to the weights can a model be specialised via fine-tuning on your own data, your own technical language or your own output format, in such a way that the result belongs to the company. With methods such as LoRA, this is possible in a manageable time even without a data centre.

A real-world example: the Llama series and its consequences

The defining example of open weights is Meta's Llama. The first version was only released to researchers in February 2023 but appeared online as a torrent within a week. Meta did not respond by shutting it off but with Llama 2 in July 2023, which was officially released for download and commercial use. With Llama 3 (April 2024) and the subsequent versions, the series became the most widely used foundation for corporate deployments and for thousands of derived models on Hugging Face.

Two developments have changed the field since then. First, the centre of open-weights development has shifted to China: DeepSeek (whose model R1 appeared under the MIT licence in January 2025 and briefly caused the share prices of Western AI suppliers to collapse), Alibaba's Qwen family, Moonshot's Kimi and Zhipu's GLM set the standards for open models today. Second, in August 2025 OpenAI published weights under the Apache 2.0 licence with gpt-oss for the first time since GPT-2, a clear sign that open weights are taken seriously as a distribution channel for developer loyalty.

For European companies, this has a practical consequence: the most powerful open models now come predominantly from China. The model itself can be run in the EU without any problems, because the weights are simply files and no data flows back. The question of origin concerns the training data, possible built-in content filters and the licence terms rather than data protection during operation.

Distinction and typical misconceptions

"Open weights means free of charge." The weights cost nothing; operation does. A model with 70 billion parameters needs several data-centre-class graphics cards for smooth inference, or quantisation, which costs quality. The total cost of ownership is made up of hardware, electricity, staff for operation and updates, and the effort for security and monitoring.

"Open weights means transparent." Without training data, nobody knows exactly what the model has learned. You can test its behaviour, but you cannot check its origin. Anyone who needs traceability, for example for regulated applications, must look for models with documented training data or set up their own evaluations.

"Open weights means secure." A locally operated model protects data from the provider, not from your own mistakes. Prompt injection, data leakage via insecure tool connections and missing access rights are just as much an issue with self-hosted models as with APIs, except that responsibility for them now lies entirely in-house.

"Open weights are weaker." The gap to the best closed models is real, but it has shrunk to a few months. For most business tasks (summarising, classifying, extracting, standard code generation), a current open model is sufficient. The lead of closed models shows above all in long agentic tasks and very demanding reasoning.

Regulatory framework: the EU AI Act

The European AI Regulation deals with general-purpose AI models in a separate chapter. For models published under a free and open licence whose parameters, architecture information and usage notes are publicly accessible, Article 53 provides for relief from documentation and transparency obligations. However, this exemption does not apply to models with systemic risk, i.e. the largest models above a certain threshold of computing power used in training. Regardless of this, companies that merely use an open model are subject to the obligations arising from the respective application, for example for high-risk systems.

Outlook

Three developments are emerging. First, open-weights models are becoming ever larger and at the same time more efficient thanks to mixture-of-experts architectures, so that models with more than a trillion parameters activate only a fraction of them per request. Second, the offering is becoming more specialised: alongside general-purpose models, open models are emerging for code, for voice agents, for image and video understanding and for individual languages. Third, the European infrastructure for operating open models is growing, from hosting providers with inference APIs through managed Kubernetes offerings to initiatives for sovereign AI cloud platforms. For mid-sized companies, the question is thus becoming less "open or closed" than "which task runs where", and a well-built system can combine both worlds.

Frequently asked questions

Is an open-weights model the same as an open-source model?
No. Open weights means that the trained parameters are available for download. Open-source AI according to the definition of the Open Source Initiative additionally requires open training code and sufficient information about the training data under free licences. Most well-known open models are open weights but not open source.

May I use an open-weights model commercially?
That depends on the licence. Models under Apache 2.0 or MIT, yes, without restriction. Community licences such as Llama's allow it with conditions. Research-only licences prohibit it. The licence is stated in the model card on Hugging Face and should be read before the first test.

Which hardware do I need to run an open model?
Small models with 7 to 14 billion parameters run quantised on a single consumer graphics card or a current notebook. Models with 70 billion parameters need one or more data centre GPUs with 80 GB of memory. The very large models with several hundred billion parameters require a GPU cluster or a specialised inference hosting provider.

Are Chinese open-weights models a data protection problem in the EU?
Not when operated locally, because the weights are files with no return channel to the manufacturer. It is different if the model is used via the manufacturer's API; then processing takes place in a third country and the usual GDPR requirements apply. Regardless of this, the licence and the model's behaviour still need to be checked.

When is an open model worthwhile compared with an API?
With high, even volumes, with strict confidentiality requirements, when fine-tuning is needed and when version stability is important. With low volumes, for tasks at the performance limit and without in-house operational know-how, the API is usually the better choice.