AI Service Providers and Models

OpenAI

Creating an API key is thankfully quite easy and the cost to use OpenAI models is affordable, even for personal use. To get started, follow the instructions on this page.

Ask ChatGPT or Microsoft Copilot How do I create an API key for OpenAI? to get detailed instructions.

The above page allows you to sign up. Then follow the instructions to create an API key.
Currently, the recommended model is GPT-5 Mini, which offers excellent results at a very affordable price.

See the OpenAI Models Overview for available models, costs, and rate limits.

Open the AutoTagger configuration dialog box via Edit > Preferences > AutoTagger and select OpenAI as the AI. The model drop-down now lists all OpenAI models supported by AutoTagger. Create an API key for the model(s) you want to use.

Don't forget to configure your rate limits for OpenAI, which depend on your usage tier. See rate limits for more information.

OpenAI frequently releases new models. We keep you informed via the IMatch release notes when we add support for new models to AutoTagger.

OpenAI offers a cost-effective solution, even for personal use, and it works on any computer, without requiring a powerful graphics card or other special hardware.

You can deposit $10 into your account and disable automatic renewal. This means when your balance is depleted, OpenAI will stop processing requests, and AutoTagger will display error messages. By disabling automatic renewal, you have full control over your spending.

With just $10, you can automatically add descriptions and keywords to thousands of images.

If you already have a 'pro' ChatGPT account, note that you'll still need to create a developer account in order to get an API key for IMatch to use. OpenAI keeps these separate.

Mistral AI

Creating an API key is thankfully quite easy and the cost to use Mistral models is affordable, even for personal use. To get started, follow the instructions on this page to create an account.

Ask ChatGPT or Microsoft Copilot How do I create an API key for Mistral? to get detailed instructions.

When you have an account, you can subscribe to a plan. For initial testing, subscribe to the free plan. No credit card needed.

The free plan allows you to create an API key, which you can then use with AutoTagger. The free plan has (obviously) rather severe rate limits (requests that can be made per day and tokens used), but it works well enough to try out the results you can expect from Mistral.

Open the AutoTagger configuration dialog box via Edit > Preferences > AutoTagger and select Mistral AI as the AI. The model drop-down now lists all Mistral models supported by AutoTagger. Create an API key for the model(s) you want to use.

Mistral offers a cost-effective solution, even for personal use. And it works on any computer, without requiring a powerful graphics card or other special hardware.

AutoTagger supports the current version of the Medium and Large models via the latest model id. When Mistral releases new model versions, AutoTagger automatically uses the new models.

Privacy

Mistral, as a European company based in France, adheres to the strict European data privacy rules and regulations. This can make Mistral the preferred choice for users—especially corporate users—who must follow the same rules or who consider privacy important.

Google Gemini

To get started, follow the instructions to create an account and get a Gemini API key. The blue button starts the process.

Just sign in with your Google account (or create one) and follow the instructions. Create a project and you get the API key you need to enter in IMatch AutoTagger. Ignore all the other information about "setting up your API key". AutoTagger takes care of all of that.

After getting your API key, select the Gemini AI in AutoTagger, add your API key, and you're ready to go:

Google Gemini excels at anything related to location (places, landmarks, tourist spots, well-known buildings), animal breeds, OCR, and many other things.

If you autotag typical vacation and travel photos frequently, give the Gemini AI a test drive. Also, if you are interested in bird photography, cars, planes, boats, and similar, Gemini AI with the Flash Light model (very affordable) should produce outstanding results.

Search for Gemini prompting tips to get some ideas. The default prompts shipped with IMatch work well for a wide range of motives. But you can of course specialize your prompts to better match your preferred style and detail level.

Ollama
(Local AI)

IMatch Learning Center

There is a free video tutorial showing how to install and use Ollama.

Ollama is an open-source and free project that enables you to run artificial intelligence (AI) applications locally on your PC, without relying on cloud connectivity. Since your images never leave your computer, there are no privacy issues.

With a moderately powerful graphics card (NVIDIA preferred) and at least 6 GB VRAM, Ollama can run smaller AI models like Gemma 4 4b or Qwwen 3.5 4b efficiently. For running larger models, more VRAM (12 or more GB) is required.

1. Installing Ollama

To get started, visit the official Ollama website and download the installation package for Windows. Double-click the downloaded file to begin the installation process, which requires no administrator privileges.

After installation, Ollama will appear in the Windows Taskbar and automatically start with Windows. This ensures that you can access Ollama quickly and easily.

2. Downloading a Model

As explained in the AutoTagger Prompting help topic, Ollama requires a model to function. Fortunately, many free models are available for various applications, including image analysis, coding, and chat.

For an overview of all available Ollama models, visit Ollama's Model Library page. For IMatch, only multimodal (vision-enabled) models are relevant.

Installing the Google Gemma 4 Model

This is a multimodal model released by Google in April 2026. It offers top-of-the-line vision processing, even for models reduced to 9b or 4b parameters. It supports 140 languages so you might get sensible and grammatically-correct descriptions, headlines and traits in languages other than English.

The 4b variant needs about 5 GB RAM on the graphics card for optimal performance. For the 9b model, we recommend a graphics card with 12 GB RAM.

• Open a command prompt window by pressing Windows key + R
• Type the word cmd and press Enter
• Type ollama run gemma4:e4b (or ollama run gemma4:12b if your graphic card has 12 or more GB VRAM) at the command prompt and press Enter

The model will begin downloading, which can take a while, depending on your internet speed. The gemma4:e4b model is approximately 10 GB in size.

Once the download completes, you're ready to start using Ollama with IMatch AutoTagger. Continue with the AutoTagger help topic to learn how to add descriptions, keywords, and traits to your images automatically.

After the model has been downloaded, you can close the command prompt window.

AutoTagger has configurations for both the 4b and 12b variants and you can just create a new setting with the model of your choice:

If you find the results of Gemma 3 lacking for your type of photography, try one of the other vision-enabled models available for Ollama.

Installing the Qwen 3 Model

Qwen is a modern model released in 2026. We recommend giving it a try and see how well it works with your images.

ollama run qwen3.5:9b

ollama run qwen3.5:4b

Note: These are thinking/reasoning models, which take longer to produce better answers. According to the Ollama and LM Studio documentation, appending the term /no_think to your prompt reduces or eliminates thinking.

Ollama Settings in AutoTagger

If you have downloaded multiple models, you can select any of them in AutoTagger:

This list shows all models AutoTagger supports, not just the models you have downloaded.

LM Studio
(Local AI)

Unlike Ollama, LM Studio has a comfortable user interface and is designed to be used both as an AI runner (like Ollama) and as an interactive tool to chat with the models installed in LM Studio.

LM Studio is 'better' than Ollama when you also want to use AI models outside of IMatch.

To download and install LM Studio on your PC, follow the easy-to-follow instructions on the LM Studio website.

Here is the LM Studio documentation, which explains all features and usage in detail.

Downloading Models

LM Studio has a comfortable model management feature. Click on the Model Search button on the left and then search for Gemma 4 or Qwen 3.5 (or whatever model is current at the time you're reading this).

Start the LM Studio Server

This is the important part. The LM Studio server makes models accessible for other applications, like IMatch.

At the bottom of the LM Studio app window, click on the Power User button:

Now click on the Developer button on the left:

Enable the server with the switch (1) and then click (2) and select the Gemma 3 model you have downloaded. Use the default settings for everything.

Using LM Studio and Models in IMatch AutoTagger

AutoTagger is already configured to support LM Studio and suitable models. Not all models you can use with LM Studio are multimodal (vision-enabled). For AutoTagger, only vision-enabled models are usable, such as Gemma 3, LLaVA, or Llama Vision.

However, you can install any other model offered by LM Studio and use it for chatting, spelling and grammar checks, text generation, SEO, math, etc. The LM community makes suggestions regarding which models to use for each purpose. Things are changing rapidly, and new models are released frequently.

Vision-enabled models are marked with this icon:

Open Edit menu > Preferences > AutoTagger and switch the AI to LM Studio. In the Model drop-down, select the Gemma 3 model available in LM Studio. We have installed the 12B Gemma 3 model in this case:

Check the prompt and any settings you need, then confirm the configuration with OK.

Select one or more images in a File Window and start AutoTagger with F7. Select the LM Studio AI and the setting you want to use and run it:

Make LM Studio Run Automatically

Click on the Gear icon button at the bottom right of the LM Studio app window to open the settings. Scroll down until you see this option and enable it:

You can now use LM Studio models in AutoTagger without starting LM Studio manually. LM Studio automatically loads the model requested by AutoTagger.

Ollama or LM Studio for Local AI?

Ollama just works mostly in the background and is the simplest solution. Use it when you only want to work with AI models within IMatch.

LM Studio allows you to chat and interact with locally installed AI models much like you would use cloud-based AI such as ChatGPT, Mistral, Copilot, or Gemini. This enables you to access AI capabilities and functions without any privacy issues or cost.

This is AutoTagger's most flexible AI connector. It allows you to connect AutoTagger with any local or remote AI service of your choice, provided that it supports an OpenAI-compatible endpoint like .../chat/completion and structured responses. Many of the tools out there do, like for example Unsloth Desktop or llama-server. If you use such tools already, you can use them with AutoTagger too.

When you select this service type for an AutoTagger setting, you configure these additional settings:

When To Use This?

If you want to run AI models with a service not directly supported by IMatch, like Unsloth Deskop or similar, this is the way to make it work with AutoTagger.

Usage Tips for llama-server

The main goal of llama.cpp is to enable running local AI with minimal setup and state-of-the-art performance on a wide range of hardware - locally and in the cloud.

llama.cpp is the core on which many AI solutions are built today. It is available on all major operating systems and supports a wide range of processors and graphics cards.

The relevant component of llama.cpp with respect to IMatch AutoTagger is llama-server. This server allows you to run any AI model supported by llama.cpp and provides an interface that IMatch can directly access via AutoTagger, similar to Ollama and LM Studio.

Using llama.cpp to run AI models is the solution with the least overhead, the best performance, and does not require any other 3rd party software. It's also something that requires some deeper computer knowledge, familiarity with command line applications etc.

Installation

Installation is simple. You download the latest release matching your operation system and graphic card from llama.cpp's GitHib repository.

For most users, the Windows x64 (Vulkan) version fits. It supports both CPU and a wide range of graphic cards. Download the ZIP file, extract the contents into a folder, and you're good to go.

You can install it also via the winget command in Windows:
winget install llama.cpp
Windows downloads and installs it, and adds it to the PATH so you can run it from everywhere. And Winget will keep llama.cpp up-to-date automatically. With winget uninstall llama.cpp you can uninstall it cleanly.

For users with a NVIDIA Graphic Card

If your PC is equipped with a NVIDIA graphics card and you want the best possible performance, use one of the Windows x64 (CUDA 12) - CUDA 12.4 DLLs (older NVIDIA graphic cards) or Windows x64 (CUDA 13) - CUDA 13.3 DLLs (more recent NVIDIA cards) versions. Download both ZIP files (llama and the CUDA DLLs) , extract them into a folder, and you're good to go.
There are also versions from AMD and Intel to get the best performance out of their graphic cards.

Models

llama-server can use models in GGUF format you have already downloaded, or download models on first use from Hugging Face using the -hf command line argument. See documentation for more information.

LM Studio uses models in GGUF format. If you have downloaded models using LM Studio, you can use the same models with llama-server.

Running llama-server

This is both the easy and, initially, complex part:

1. Open a command prompt window
2. cd into the folder where you unpacked the llama.cpp binaries
3. Assuming you have downloaded your model already and placed it into C:\data\ai-models;
   llama-server -m "C:\data\ai-models\gemma-4-12B-it-Q4_K_M.gguf" --mmproj "C:\data\ai-models\mmproj-gemma-4-12B-it-BF16.gguf" --port 11234 -c 8192 -ngl 99 -t 8 --parallel 2 --temp 1.0 --top-p 0.95 --top-k 99 --flash-attn on --threads-draft 4 --reasoning-format none --reasoning off
   This starts llama-server with the Gemma 4 12b model. It works the same with other models. Note that vision-enabled models always use a model and a so-called projector model (with the --mmproj parameter).

   Set the -t parameter to the number of physical processors in your computer.
   The --parallel n controls how many parallel requests llama-server supports. Experiment with this value (the default is 4), and set the threads option in the AutoTagger settings to the same value you have used for --parallel. This tells AutoTagger how many parallel requests it is allowed to make.

   Finding a good setting for -parallel for your CPU and GPU can make a big performance difference. Start with --parallel 2 and 2 threads in the AutoTagger settings. Run AutoTagger for a set of 20 or 30 files and measure how long it takes to process them. Increase --parallel and threads and repeat the test. If it is faster, increase a again. If it is slower or AutoTagger reports errors, decrease both values again. Keep an eye on the Performance tab in Windows Task Manager to see how the CPU and GPU is utilized.

This screen shot shows standard settings for using a running instance of llama-server from AutoTagger. AutoTagger will use whatever model you have loaded in llama-server. Note that the port number (11234) may be different for your installation.

This is a good starting point. All llama-server parameters are explained in their documentation and you can always ask an AI for details.

If you load a different model into llama-server while IMatch is running, you need to close and re-open the database. IMatch retrieves and caches the model list only once, during the first request for performance reasons.

Usage Tips for Unsloth Desktop

Unsloth Desktop is a free, open-source app for running and training AI models on your own local hardware. The company behind Unsloth is also known for creating high-quality distilled / quantized models from open weight models and making them available on Hugging Face.

To install Unsloth Desktop, follow the instructions on their website. To use Unsloth like Ollama or LM Studio from AutoTagger, you run it as an api endpoint, see https://unsloth.ai/docs/basics/api. Unsloth then runs as a server, and AutoTagger can connect to it as it would connect to Ollama or LM SAtudio.

One advantage of Unsloth Desktop over Ollama or LM Studio is that it can also perform tasks like local training, creating images, video, audio and much more.

Another advantage is that Unsloth Desktop is aware of the latest models Unsloth provides, an they often work 'better' or 'faster' than other models available on Hugging Face.

This screen shot shows some typical settings for using a running Unsloth Desktop instance from AutoTagger. Note that the port number (8888) may be different for your installation.

Unsloth has an option that allows to specify the model when making requests. If this is enabled, you can specify the model name in your AutoTagger settings as shown above.

Thinking Models

Thinking models like Google Gemma 3, Qwen 3.6 etc. may produce better results - but they also take a lot more to respond.

In our experiments, we did not notice much of a difference between the quality of descriptions and keywords produced in thinking mode or in non-thinking mode. While thinking mode can be very helpful to solve complex programming or math tasks, it does not as much for simpler tasks like describing images.

Since Unsloth Desktop uses llama-server internally and allows to provide additional llama-server parameters in the model settings, you can disable thinking by providing these twp parameters:

--reasoning-format none --reasoning off

You may also experiment, e.g.:

--reasoning-format none --reasoning auto --reasoning-effort minimal

where --reasoning-effort supports minimal, low, medium, high, xhigh or max (default: default)

As so often, what works best for you will require some testing with various settings and checking the returned descriptions, keywords, and traits.

In our tests on a NVIDIA RTX5080, disabling thinking reduced the response time from about 5 seconds to 1 second per image, for a complex prompt that returns a description, hierarchical keywords and a headline.

See the llama-server documentation for additional parameters and their meaning.

Things change fast in AI. We wrote this in August 2026. At the time you read this, other parameters or tokens you can include in your prompt may be more effective. Always refer to the current documentation of Unsloth Desktop and llama-server.

The site https://canitrun.dev/ allows you to select your graphic card and input the amount of RAM in your computer and then shows which models you can run locally. A good starting point for your own experiments with local AI and llama-server or Unsloth Desktop.