FileMaker 2025 integrates LoRA training - AI fine-tuning directly from the database for the first time

With the new version of FileMaker 2025, Claris introduces a function that was previously reserved for the world of AI research: the fine-tuning of large language models (LoRA) - directly from the database, locally and without a cloud connection.

Developer and author Markus Schall has examined this approach in a comprehensive technical article on the gofilemaker website and summarized it in a new post on his blog. His conclusion: FileMaker makes AI training ready for production - stable, comprehensible and secure.

What is FileMaker anyway - and why does it suddenly play a role in AI?

FileMaker has been a tried-and-tested platform for developing individual business solutions for decades. Originally started as a database system, FileMaker has evolved into a versatile tool that combines data management, process automation and user interfaces in a single environment. Companies use FileMaker to digitize internal processes - for example for merchandise management, project management, CRM or accounting. The big advantage: applications can be created and customized without in-depth programming knowledge.

Claris, a subsidiary of Apple, has gradually developed FileMaker into a modern low-code platform in recent years. Today, it can be used not only to process traditional business data, but also to integrate AI functions - such as text analysis, automation and now new: training your own language models. This makes FileMaker a kind of bridge between traditional business software and the world of artificial intelligence - simple, comprehensible and with the familiar Mac quality.

Why FileMaker is interesting for companies

While many AI platforms are highly specialized or purely cloud-based, FileMaker relies on local processing and simple integration into existing processes. Companies that do not want to transfer their data to external providers will find an alternative here that offers technical and organizational security. Applications created in FileMaker run on Mac, Windows and the web - and can be flexibly adapted to different industries.

Medium-sized companies in particular appreciate FileMaker because it allows them to develop their own solutions without getting bogged down in complex software projects. The platform offers graphical tools, tables, layouts, scripts and automation - all from a single source.

The new AI extension in FileMaker 2025 now opens up another area: training and customizing your own AI models, without external infrastructure, without data transfer and without complicated command lines.

LoRA training as a key technology of modern AI

Fine-tuning with LoRA (Low-Rank Adaptation) is considered one of the most important advances in the development of language models. Instead of completely retraining a neural network, only small, specialized additional layers - so-called adapters - are trained. This saves computing time and energy and makes it possible to adapt a basic model to specific topics, tonalities or specialist languages.

Until now, however, this process has been highly technical and error-prone. Most users worked with complex Python frameworks such as PEFT, Axolotl or kohya_ss and had to fight their way through a multitude of parameters, file formats and GPU dependencies. Each training environment was its own experiment - stable if you were lucky, but rarely reproducible.

The new approach: LoRA training via script

With FileMaker 2025, LoRA is integrated and automatable for the first time. A new script command - "Fine-Tune Model" - allows you to train your own language models directly from a FileMaker database. The process runs completely locally via the AI Model Server, which uses the new MLX framework on Apple Silicon systems (M1, M2, M3).

This allows companies or developers to adapt their AI models to company-specific data - such as support dialogs, product descriptions or internal documentation.

The training uses either data from FileMaker tables or JSONL files with a chat structure and creates a new, locally usable model with the prefix fm-mlx-....

Advantages of this approach:

* No terminal, no dependencies, no cloud constraints.
* Data remains completely local on your own system
* Reproducible results thanks to clearly defined parameters.

Ideal for companies that want to use AI securely and traceably.

Two worlds in comparison - classic PEFT vs. MLX in FileMaker

Markus Schall compares the two approaches in detail in the detailed technical article. While classic tools such as Axolotl and kohya_ss open up almost every adjustment screw, FileMaker deliberately focuses on stability and ease of use. Here, fine-tuning becomes a process that can be integrated into everyday work - no longer an experiment in the terminal window.

FileMaker brings order to an area that for a long time was only accessible to specialists", explains Markus Schall. "What used to take a whole day of troubleshooting, GPU drivers and JSON conversions is now a script command. This is not a step backwards, but the step that AI needs to take to become part of everyday life."

The article on gofilemaker describes the technical process, shows example scripts, JSONL training files and explains the internal differences between the MLX-based FileMaker models and classic PEFT LoRAs.

A tool instead of an experiment

For companies in particular, this opens up a new way of adapting AI models in a targeted, secure and data protection-compliant manner. As the entire training takes place locally, there is no need to transfer sensitive data to external services - a crucial point in industries where data protection, documentation and traceability are key requirements.

FileMaker 2025 thus takes LoRA training out of the experimental phase.
A complicated command line exercise becomes a reproducible process - documentable, script-controlled and in the familiar FileMaker structure.

The change is reminiscent of the transition from handwritten code to visual database applications: the same functionality, but much more accessible.

Outlook: Local AI systems as the standard

In his analysis, Markus Schall emphasizes that this is just the beginning of a larger development. MLX and Apple Silicon are currently giving rise to entire ecosystems of local AI applications - from language models to vector databases and image generators. FileMaker 2025 fits in here as an interface between productivity, data storage and AI logic.

"The decisive difference is not technical, but structural," says Schall. "FileMaker brings the fine-tuning to where the data is located anyway - in the business process itself."

About the author

Markus Schall is a software developer, system consultant and author. With his company goFileMaker, he develops ERP solutions based on FileMaker and supports companies in the digitalization of internal processes. At the same time, he runs the publishing house M. Schall Verlag and publishes books on topics such as digitalization, personal development and AI in everyday life.

Frequently asked questions

* What exactly is FileMaker - and why is it mentioned in connection with AI?
FileMaker is a software platform from Claris (Apple) that enables companies to develop their own applications for organization, administration and digitalization - without traditional programming. FileMaker can be used to map business processes, build databases, set up automation and, since version 2025, also integrate AI functions. The new LoRA function now extends FileMaker with the option of training language models directly in your own system.
* Do I have to be a programmer to start AI training in FileMaker?
No. That is one of the main advantages. In FileMaker, everything runs via clearly structured dialogs or simple script commands. The LoRA training can be started via a prepared button or a short script - without any command line or additional software. The technical background remains hidden in the system, control is via familiar FileMaker interfaces.
* How does my company benefit from training its own language model?
Your own model can be tailored to the language, terms and processes of your company. As a result, it reacts more precisely, understands technical terms correctly and delivers better results in internal processes - for example in customer communication, support or automated text output. Instead of using a "generic" AI model, train one that speaks your language.
* Does the AI training run in the cloud or on my own devices?
LoRA training in FileMaker 2025 runs entirely locally on Apple Silicon machines (e.g. MacBook, Mac mini, iMac with M1, M2 or M3). Your data does not leave your company. This is particularly interesting for industries in which data protection and internal confidentiality are important - for example, service providers, doctors, law firms or craft businesses.
* How time-consuming is it to train such an AI model with FileMaker?
Much easier than before. Traditional AI training usually requires several tools, technical knowledge and GPU infrastructure. In FileMaker, it is sufficient to provide existing data records (e.g. customer inquiries or product texts) and execute a script command. Training takes between a few minutes and a few hours, depending on the amount of data, but runs automatically in the background.
* Is this more of an experiment or is it already intended for practical use?
This is already intended for practical use. FileMaker 2025 makes fine-tuning reproducible and stable. It is not a test function, but a new, integrated tool that can be used productively in the FileMaker environment. Claris has designed the function to run reliably - even in corporate environments with clear IT structures.
* Can I also use the trained model outside FileMaker at a later date?
The trained model is currently available in Apple MLX format, which is specially optimized for FileMaker's AI Model Server. In the future, there will probably be export options, for example to open formats such as GGUF or ONNX. However, the model can already be used completely within FileMaker - e.g. in your own layouts, chat functions or automations.
* What distinguishes FileMaker from other AI platforms?
Many AI platforms are geared towards developers or cloud services. FileMaker, on the other hand, combines database, user interface and AI functionality in one system - locally, comprehensibly and without subscription dependencies. For companies, this means full control over data, processes and results. AI does not become a third-party system, but part of the company's own infrastructure.

gofilemaker.de - MSITS
Hackenweg 97
26127 Oldenburg
Germany

https://gofilemaker.de/en/
Herr Markus Schall
+49 (0) 4 41-30 43 76 40
+49 (0) 4 41-30 42 51 32
info@gofilemaker.de

Development of customizable ERP software with CRM, billing and merchandise management for macOS, Windows and iOS based on the Claris FileMaker platform. Design of local AI systems for connecting databases to local LLM language models. Development and distribution of database systems, ERP software and merchandise management systems based on the FileMaker platform. Experience with FileMaker databases since 1994, winner of an FMM Award 2011, awarded by FileMaker Magazine.

This release was published on openPR.