Local AI with Ollama and LM Studio: language models on your own machine

This article is an overview based on vendor information, documentation and publicly available reporting, not a scored test of our own. Prices and features change quickly; the vendor's own terms always take precedence. We only score what we have used ourselves for weeks: those tools are listed under AI tools compared.

Two tools, one principle

Both programs solve the same problem: an open language model is downloaded as a file and then executed entirely on your own computer. There is no account, no queue and no connection to a vendor once the model sits on your disk. Under the bonnet both rely on the open-source library llama.cpp, and LM Studio additionally uses Apple's MLX framework on Macs.

Ollama is a command-line tool under the MIT licence for macOS, Windows and Linux. One command downloads a model, a second starts the conversation, and in the background a local server runs whose interface is compatible with the OpenAI format. Applications that previously talked to ChatGPT through the API can be switched to a local model with a few changes. Since 2025 there is also a desktop app with a chat window into which you can drag text files and PDFs.

LM Studio starts from the screen: a graphical interface with a model catalogue, chat window and settings, where you load models from Hugging Face with a click and try them straight away. It too provides a local server, according to the vendor in both the OpenAI and the Anthropic format. If you do not write code, LM Studio gets you going faster; if you want to automate workflows, Ollama usually does.

Which models are worth considering

The catalogue of both tools covers the familiar open model families: Llama from Meta, Mistral Small from Mistral AI, Gemma from Google, Qwen from Alibaba and DeepSeek. All come in several sizes, stated in billions of parameters, and the size determines almost everything: memory footprint, speed and quality.

For orientation, as of September 2026 and according to the Ollama catalogue: Qwen3 ranges from a variant with 0.6 billion parameters up to 235 billion; the 8B version takes around 5 GB on disk, the 14B version around 9 GB. Mistral Small with 24 billion parameters needs about 14 GB, Gemma 4 comes in steps from roughly 2 to 31 billion parameters, and Llama 3.3 with 70 billion parameters runs to a good 40 GB.

A note on DeepSeek: the small variants of DeepSeek-R1 that run on laptops are distilled models built on Qwen or Llama. The actual model with 671 billion parameters occupies around 400 GB and is irrelevant for ordinary office machines.

Hardware: memory, graphics card, quantisation

The decisive figure is the memory the model has to fit into: the video memory (VRAM) of an Nvidia or AMD card, or on Apple machines with an M-series chip the unified memory shared by CPU and GPU. If the model fits, it answers fluently; if it does not, the processor has to help out and seconds become minutes.

To make models fit on everyday machines at all, they are quantised, that is, stored with reduced numerical precision. The common 4-bit quantisation needs roughly a quarter of the memory of the original version, with a small loss of quality. As a rule of thumb: a model with 7 to 8 billion parameters then takes around 5 GB, one with 14 billion around 9 GB, one with around 30 billion around 20 GB, and one with 70 billion more than 40 GB. On top of that comes memory for the context, meaning long documents in the conversation.

According to the vendor, LM Studio recommends at least 16 GB of RAM and, on Windows, a graphics card with at least 4 GB of VRAM; Intel Macs are no longer supported. In practice that means: an office laptop with 16 GB runs models up to about 8 billion parameters usably, a Mac with 32 or 64 GB of unified memory or a PC with a 24 GB graphics card reaches the class around 30 billion. Beyond that, workstation hardware begins.

What it is worth against ChatGPT and Claude

Honestly: considerably less as soon as the task gets demanding. The models that run on a laptop are orders of magnitude smaller than the models behind ChatGPT or Claude. For multi-step analysis, long context, complex code or tasks with many constraints the gap is clearly noticeable in daily use, and small models invent facts more often.

For a clearly bounded share of office work it is still enough: summarising texts, pre-sorting emails, filing documents into categories, turning bullet points into a first draft, answering questions about an uploaded PDF. Models such as Mistral Small or Gemma are, according to their vendors, trained on German; the quality is decent but does not reach the level of the large cloud models we set side by side in our comparison of AI assistants.

The second price is speed. On suitable hardware small models answer quickly; on tight hardware they are markedly slower than any cloud service. If you need both, local processing and top quality, you have to invest in hardware or split the tasks.

Cost and data protection: what running locally solves, and what it does not

The software costs nothing. Ollama is under the MIT licence, and LM Studio has, according to the vendor, been free for use at work since July 2025; only resale and operating it as a service for third parties are excluded. Both vendors now offer optional paid services, Ollama for instance cloud models from around 20 US dollars a month, LM Studio an enterprise tier with single sign-on. For purely local use you need neither. The real costs sit in the hardware and in the time of the person who sets it all up and maintains it.

The data protection advantage is real: when running locally, inputs and documents do not leave the machine, and both vendors say so explicitly. There is no processor and therefore no data processing agreement, no question about the processing location and no training on your data. That is the same reason a locally run Stable Diffusion is the cleanest solution for images involving personal data. Just make sure that cloud features such as Ollama's "-cloud" models stay switched off, otherwise the advantage is gone.

Two things running locally does not solve. First, responsibility: the processing still belongs in the record of processing activities, access rights and deletion periods have to be regulated internally, and the AI Act's literacy obligation applies regardless of where the model does its computing. Second, model licences: Mistral, Qwen and DeepSeek are under Apache 2.0 or MIT licences, Llama under Meta's community licence with an attribution requirement and a user threshold, Gemma under Google's terms with a list of prohibited uses. For products built on a model, that needs checking. The remaining checkpoints are in our guide to AI tools and the GDPR.

Who local AI is worth it for

Anyone who works with data that must not leave the machine: law firms, medical practices, tax advisers, HR departments, research with confidential material. Developer teams building an AI feature as a prototype before choosing a vendor. And businesses with many uniform tasks where a cloud service's bill grows quickly, such as sorting or summarising thousands of documents.

It is not worth it for anyone who above all needs top quality, has no suitable hardware, or has nobody in house to select and update models and answer questions. If you want European processing without running anything yourself, a paid plan of Le Chat from Mistral is often a better fit than an underpowered laptop. A mix is frequently the sensible choice: confidential material locally, demanding work without personal data in the cloud.

In short: Ollama for the command line and automation, LM Studio for getting started through a graphical interface, both free. From 16 GB of RAM small models run usably; for more you need video memory. Quality well below ChatGPT and Claude, but no data leaves the machine.