Compared · checked 2026-10-24

The best free apps to run AI chat on your own computer

A local AI model answers on your own hardware, so your chats do not go to a cloud service. Some of these apps install like any program, two run in the browser. Here is what each one offers, from its own pages.

We make Assistant, one of the tools on this page. Every fact about the other tools comes from their own pages, linked under each one, as they read on 2026-10-24. Prices and features change: check each site before you choose.

How we compared them

• Free to run models locally
• What you need: an install, a GPU, a given OS or browser
• Works offline once a model is downloaded
• Model choice and size
• Extras: chat with documents, images, a local API

At a glance

1. Assistant (ours)

Best for: Checking what your PC can run and chatting with a small model in the browser, with no install, then using it across ITpris Verktyg apps.

Good

• "Can my PC run it?" tests WebGPU and memory and rates each model before you download it
• Asks before each download, then the model stays on your device
• Builds small one-file apps in a sandboxed preview, and can open and drive other ITpris Verktyg apps
• Chats are kept on your device and export as Markdown or JSON
• The AI button in the other ITpris Verktyg apps uses the same models, and can also use Ollama or LM Studio on your computer, or your own key for a cloud model
• Works on your phone, tablet and computer in the browser, even offline

Not so good

• Small models only (up to 7B), so answers are weaker than the bigger models desktop apps can run
• Text only: images cannot be sent to the model
• No local API for other programs
• Needs WebGPU, and it is still in beta

Open Assistant

2. LM Studio

Best for: Running bigger open models in a free desktop app, with a chat window and a local API.

Good

• Runs llama.cpp models, and MLX models on Apple Silicon
• Chat with documents and MCP servers
• Serves models on OpenAI-like endpoints
• Can operate entirely offline

Not so good

• 16 GB of RAM recommended, and 4 GB of dedicated VRAM on Windows
• Intel Macs are not supported
• Paid plans ($20 and $100 a month) are for its cloud models

LM Studio website · Sources: LM Studio pricing, LM Studio system requirements, LM Studio docs

3. Ollama

Best for: Developers who want local models behind an API for their own apps and coding tools.

Good

• Local API, with OpenAI and Anthropic compatible clients
• NVIDIA and AMD Radeon GPU support on Windows
• Vision models that take images
• MIT licence

Not so good

• Built around the command line and the API
• Its paid plans ($20 and $100 a month) are for cloud models and usage credits
• On Intel Macs it runs on the CPU only

Ollama website · Sources: Ollama pricing, Ollama on Windows, Ollama on macOS, Ollama vision, Ollama OpenAI compatibility, Ollama on GitHub, Ollama docs, Ollama quickstart

4. GPT4All

Best for: A private desktop chatbot that runs on ordinary laptops, even without a GPU, and chats with your files.

Good

• No API calls or GPUs required, per its README
• LocalDocs: chat with your own documents
• Says no data leaves your machine

Not so good

• The Linux build is x86-64 only
• Needs at least an Intel Core i3 2nd Gen or AMD Bulldozer CPU

GPT4All website · Sources: GPT4All home page, GPT4All on GitHub

5. Jan

Best for: An open-source ChatGPT-style desktop app that mixes local models with cloud ones when you want.

Good

• Local OpenAI-compatible API at localhost:1337
• MCP integration
• Can also connect to OpenAI, Anthropic, Mistral and Groq

Not so good

• Its own guide: 8 GB of RAM for 3B models, 16 GB for 7B, 32 GB for 13B on Mac

Jan website · Sources: Jan home page, Jan on GitHub

6. WebLLM Chat

Best for: Trying local AI in the browser with no install, including vision models.

Good

• Runs in the browser with WebGPU, with no server-side processing
• Vision model support: send images
• Works offline after the initial setup

Not so good

• Needs WebGPU, so not every browser or computer can run it
• Custom models must be compiled to the MLC format

WebLLM Chat website · Sources: WebLLM Chat on GitHub

Which one to choose

Assistant is our pick for private AI chat with nothing to install: it tells you first which models your computer can run, asks before each download, keeps the model and your chats on your device, works offline after that, and can build small apps and drive the other ITpris Verktyg apps. Its models are small (up to 7B) and text only, so do not expect the answers of a big desktop model.

Choose LM Studio if you can install an app, have 16 GB of RAM and want bigger models, document chat, MCP and a local API. Choose Ollama for models behind an API for your own tools, GPT4All for older laptops without a GPU, Jan for an open-source app that can mix local and cloud models, or WebLLM Chat to send images to a model in the browser.

Questions

Is local AI really private?

The model runs on your computer, so a local chat does not have to leave it. GPT4All, Jan, LM Studio and WebLLM Chat say so on their own pages. In Assistant the chats stay on your device; it goes online to download the model you choose. Several apps can also connect to cloud models, as can the AI button in the other ITpris Verktyg apps when you add your own key; those chats do go to that provider.

What computer do I need?

LM Studio recommends 16 GB of RAM; Jan suggests 8 GB for 3B models and 16 GB for 7B on a Mac. GPT4All runs without a GPU. The browser apps need WebGPU; Assistant's system check tells you which of its models will run before you download one.

Do these work offline?

LM Studio, Jan and WebLLM Chat say they work offline once the model is downloaded. Assistant does too, after the first visit and the model download.

Which is the most powerful?

The desktop apps, such as LM Studio, run larger models than the browser apps. Assistant and WebLLM Chat trade size for having nothing to install.

Names of other products are trademarks of their owners. ITpris Verktyg is not affiliated with them.