Generative AI is usually used in everyday work via cloud services. Every request leaves the company and is processed on external servers. For organisations with strict data protection or confidentiality requirements, this is a recurring obstacle, for example with contract drafts, customer data or unpublished content.
An alternative is local models: language or image models that are run directly on your own computer. The data stays where it is created. Two current overviews show that this path is now viable even without deep specialist knowledge – with different focuses.
Apple's local model on Mac hardware
The most obvious entry point is hardware that already comes with a model. On Macs with an Apple chip, a local foundation model is available – a pre-trained base model that serves as the basis for various applications. However, it is not visible as a standalone chat app; it has to be accessed via a workaround. Anyone who knows how the access works gets a chatbot that manages without an internet connection and without a cloud account.
The practical advantage lies less in performance than in the starting position: the model is already present on suitable hardware and is specifically tuned to it. The memory requirements that otherwise make local models a problem matter less here because the system handles execution.
Tools beyond the well-known cloud services
The second perspective is broader: there is a range of applications with which AI models can be run on your own computer – from chatbots to image generation. They thus compete directly with established cloud offerings such as ChatGPT, Claude or Midjourney, without requiring their infrastructure.
The range extends from programs that primarily provide text models for dialogue and text work to tools for local image generation. What they have in common is that the model is downloaded and run on your own machine. Quality therefore depends directly on the chosen model and the available hardware.
Where the limits lie
Local models promise privacy but demand resources in return. Memory requirements are the most frequently cited bottleneck: model files take up space, and a usable response speed requires sufficient working memory and computing power. Older or poorly equipped workplace computers quickly reach their limits here.
In addition, locally runnable models are generally smaller than the large cloud variants. For summaries, drafts, rephrasing or simple code snippets, this is often sufficient. For complex tasks with long context, a quality difference is to be expected. A mixed operation is therefore realistic: sensitive content locally, non-critical tasks continue in the cloud.
Assessment for corporate use
For evaluation in a corporate context, three questions are particularly relevant:
- Which data is affected? The more sensitive the content, the greater the benefit of local processing – regardless of how well the model performs.
- Which hardware is available? Where suitable devices are already in use, no additional infrastructure costs arise. Otherwise, the hardware requirement is part of the calculation.
- Who maintains the operation? Local installations must be updated and maintained on individual workstations. This task does not apply to cloud services.
A clearly defined initial trial makes sense: one use case, one team, a defined period. Only when quality and speed convince in everyday use is expansion worthwhile. In parallel, it should be defined which data categories may in future be processed exclusively locally – otherwise the data protection advantage remains theoretical.
What this means for web and software projects
For projects with increased data protection requirements, this creates an additional option between completely doing without AI and using external services. In content creation for websites, for text variants or when working with internal documents, local models can be used where a cloud transfer is not justifiable. For architecture decisions, this means above all one thing: AI functions should be connected in such a way that the model can be swapped behind the interface – locally or in the cloud, depending on the protection requirements of the specific use case.