Open-Source AI Models: what they are, benefits, risks and opportunities

Open-source AI models are playing an increasingly important role in the artificial intelligence landscape, offering businesses new opportunities for data control, customization and management. When the Chinese laboratory DeepSeek released its R1 model in January 2025, the world discovered that it could compete with some of the most expensive artificial intelligence systems developed in Silicon Valley while reportedly being trained with a significantly lower budget than its US competitors. Moreover, the model was made freely available for download.

From that moment on, many began to wonder whether AI was following the same path as open-source software in the 1990s. In reality, the concept of openness in artificial intelligence is more complex and can have different meanings depending on the model and the organization behind it. Understanding these differences has become a strategic as well as a technical capability.

In traditional software, open source means publicly available source code that can be modified and redistributed under specific licenses. In the AI world, however, a model's openness is measured across at least four dimensions: model weights, training code, training datasets and licensing terms.

Many so-called open-source AI models release only their weights, making them open-weight models. In these cases, both the datasets used for training and the processes behind the model's development remain undisclosed.

An open-weight model can be compared to a fully furnished apartment that is ready to use. It can be modified, customized and adapted to specific needs, but users do not have access to the original blueprint or the information required to rebuild it from scratch.

This distinction is essential to understanding the ongoing debate around AI transparency. Access to model weights provides flexibility and customization opportunities, but it does not necessarily guarantee full transparency throughout the development process.

The main players in the open-source AI landscape follow very different approaches. Meta, with its Llama family of models, provides access to model weights through a community license that does not fully meet traditional open-source criteria. Training datasets and training code remain proprietary.

Mistral takes a more open approach to licensing, aiming to build a European alternative in artificial intelligence while keeping training datasets private.

DeepSeek has chosen to release model weights and part of its code under an MIT license. However, the absence of publicly available training datasets continues to fuel debate around the model's transparency and the actual efficiency of its reported costs.

Through its Qwen family, Alibaba provides access to numerous models under highly permissive licenses, creating one of the most widely adopted open-weight ecosystems worldwide. Even in this case, however, training datasets remain unavailable to the public.

Among the few organizations pursuing transparency across the entire development lifecycle is AI2, the non-profit organization behind the OLMo family of models. It publishes datasets, code and training methodologies, although performance still trails that of the leading proprietary models.

For a manufacturing company such as GEWISS, open-source AI models can represent a concrete industrial opportunity. The ability to run models on internal servers allows organizations to maintain direct control over sensitive process, product and know-how data.

For intensive use cases, proprietary infrastructure may also help reduce recurring costs compared with external AI APIs based on token-based pricing models.

Access to model weights enables advanced customization of artificial intelligence systems, allowing them to be adapted to an organization's technical language and specific business processes.

Another important advantage is the reduction of vendor lock-in, namely dependence on a single technology provider, an increasingly relevant issue in the current geopolitical environment.

The adoption of open-source AI models still requires a balanced assessment of trade-offs. According to independent benchmarking platforms such as Artificial Analysis, open models remain slightly behind the most advanced proprietary systems, although the performance gap is narrowing rapidly.

Managing infrastructure independently also requires specialized expertise, operational capabilities and dedicated hardware resources that not all organizations possess internally.

In addition, self-hosting increases responsibility for cybersecurity, maintenance, governance and compliance, without the level of support typically guaranteed by commercial vendors.

The most interesting aspect is the speed at which the market is evolving. According to analyses by Epoch AI, the performance gap between the best open models and proprietary alternatives is now measured in months rather than years.

Eighteen months after the release of DeepSeek R1, one key question remains: how much control, customization and responsibility is a company willing to take on directly, and how much would it rather delegate to technology providers?