Run AI on your own computer: 9 apps and what to check first

On your computer or hosted?

Terms in our glossary: self-hosted vs cloud-hosted, open weights vs open source, parameters and quantization. Which models are open, and under what license, is on our Open Source page.

What your computer needs

• Memory is the limit. The Ollama llama2 page says 7b models need at least 8 GB of RAM, 13b models at least 16 GB and 70b models at least 64 GB.
• A model is loaded fully into memory. The llama.cpp docs say memory and disk needs are the same, and give an example: an 8B model is 32.1 GB in its original form and 4.9 GB after 4-bit (Q4_K_M) compression. Longer conversations need more memory on top.
• Graphics memory helps, but is not required. The Ollama quickstart says that with less VRAM than a model needs, Ollama can use system RAM, though responses may be slower. Macs use unified memory for this.
• Your operating system matters. LM Studio supports only Apple silicon Macs, recommends 16 GB of RAM or more on a Mac, and on Windows recommends 16 GB of RAM and 4 GB of dedicated VRAM.

You will often see models offered in smaller, compressed versions. Quantization stores a model's numbers with less precision so the file is smaller, and the llama.cpp docs say it can cost some accuracy. GGUF is a file format used by llama.cpp, LM Studio, GPT4All and Ollama, according to Hugging Face's docs. In llama.cpp's table for an 8B model, Q4_K_M is about 4.9 bits per weight and Q8_0 about 8.5.

Desktop apps with a chat window4

Install like any other program, then pick and download a model inside the app.

• LM StudioLM Studio · Desktop appA desktop app for chatting with models that run on your computer. It has a built-in model browser that downloads from Hugging Face, plus a local API and a headless mode.License: Proprietary, free to use. Its terms grant a limited license for personal and internal business use. Not open source.Cost: The pricing page lists Free at $0 for running local models. Bionic+ is $20 a month and Pro is $100 a month, and add cloud models and higher limits.Works on: macOS on Apple silicon (macOS 14.0 or newer; Intel Macs are not supported), Windows (x64 and ARM) and Linux (x64 and ARM64).Privacy: The app privacy policy (effective June 2026) says the app has no telemetry. Update checks send the app version, OS and IP address. The docs say chat works offline once a model is downloaded; searching and downloading models needs internet.Needs: Recommends 16 GB of RAM or more on a Mac (8 GB may work with smaller models) and, on Windows, 16 GB of RAM and 4 GB of dedicated VRAM.
• JanJan (janhq) · Desktop appAn open-source desktop chat app with a built-in model Hub. It can also connect to cloud providers if you choose.License: Apache 2.0, as stated on its GitHub repository.Cost: The pages we opened state no price. The project is open source.Works on: Windows, macOS and Linux.Privacy: The docs say usage tracking is opt-in, chosen at first launch, and that Jan never tracks chat content, prompts or files. They say no internet is needed once models are downloaded. Cloud models receive your messages.Needs: The GitHub README says 8 GB of RAM for 3B models, 16 GB for 7B and 32 GB for 13B (macOS 13.6 or newer; Intel Macs are not supported).
• GPT4AllNomic · Desktop appA desktop app and Python library for running language models privately on ordinary computers. The README says no GPU is required.License: MIT, as stated in its GitHub README.Cost: The pages we opened state no price. The README calls it open source and available for commercial use.Works on: Windows, macOS (Monterey 12.6 or newer) and Linux. The README and the system requirements file disagree on ARM support, so check both.Privacy: The README describes it as private and on-device. We found no telemetry statement on the pages we opened.Needs: The system requirements file lists 16 GB of RAM as a minimum for Windows and Linux (8 GB is enough for 3B models) and a CPU that supports AVX or AVX2.The latest release we could see is v3.10.0 from February 2025, so we could not confirm the project is still actively maintained.
• AnythingLLMMintplex Labs · Desktop appAn app for chatting with your own documents. The desktop version comes with a built-in model and can also use other providers.License: MIT, as stated on its GitHub repository and home page.Cost: The home page calls it a free, private AI assistant and says the desktop app has no token limits. It does not say whether paid tiers exist.Works on: macOS, Windows and Linux desktop apps, a Docker version for self-hosting and a mobile app for Android.Privacy: The desktop privacy policy (July 2025) says anonymous usage telemetry is sent only if enabled and can be turned off in settings, and that chats and documents are not transmitted from your system. Web search is on by default through You.com and needs internet.Needs: The docs recommend 16 GB of RAM and an 8-core CPU for a basic experience, and a GPU on Windows (about 8 to 12 GB of VRAM or more is ideal).

Runners you control from a terminal4

Command-line tools that run models and often a local server that other apps can use.

• OllamaOllama · RunnerDownloads open models and runs them on your computer, from a desktop app or the terminal, and runs a local server that other apps can connect to. It can also use Ollama's cloud models.License: MIT, as stated on its GitHub repository.Cost: The pricing page lists a Free plan at $0 that includes running models locally. Cloud models need a paid plan or credits, except one that the Free plan includes. Pro is $20/mo and Max is $100/mo.Works on: macOS, Windows, Linux and Docker.Privacy: The privacy policy (updated March 2026) says prompts and responses processed locally are not collected, stored or sent to Ollama. Cloud models do send prompts to Ollama.Needs: The quickstart says its example model (about a 7.2 GB download) recommends 8 GB of available VRAM, or unified memory on a Mac. With less, Ollama can use system RAM but responses may be slower.Models come from the Ollama library with a pull command. Beginners may find the terminal harder than a desktop app.
• llama.cppggml-org · RunnerThe open-source C/C++ engine that many other apps build on; LM Studio's and Jan's docs say they run GGUF models through it. It has a command line and a server with a built-in web page, but no desktop app.License: MIT, as stated on its GitHub repository.Cost: The pages we opened state no price. It is an open-source project.Works on: Windows, Linux, macOS and Android.Privacy: We found no privacy or telemetry statement in the README or server docs. An offline flag makes the server use only local files.Needs: The README states no minimum requirements. It supports Apple silicon, x86 CPUs and NVIDIA and AMD GPUs, and can split a model between CPU and GPU.You bring your own GGUF files, or use a flag that downloads from Hugging Face. We describe it as a building block rather than a first app.
• llamafilemozilla-ai (GitHub) · RunnerPacks a model and the engine that runs it into a single file that you start on your own computer.License: Apache 2.0. Mozilla's changes to llama.cpp and whisper.cpp are MIT.Cost: The pages we opened state no price. It is an open-source project.Works on: macOS, Linux, BSD and Windows (add a .exe extension; executables over 4 GB need the model weights kept in a separate file).Privacy: The README does not mention telemetry or data collection.Needs: The README gives no RAM figures. Needs depend on the model, and larger models suit powerful hardware or GPUs.Ready-made llamafiles are downloaded from Hugging Face. It is used from the command line. The repository now lives under mozilla-ai.
• KoboldCppLostRuins (community) · RunnerA single downloadable program for running GGUF models, with a launcher window, a command line and a built-in web chat page.License: GNU AGPL v3.0 for KoboldCpp and its web page, KoboldAI Lite. Bundled parts such as GGML and llama.cpp are MIT.Cost: The README calls it free, open-source software.Works on: Windows, Linux, macOS on Apple silicon (Intel Macs need a build from source) and Android through Termux.Privacy: The README does not mention telemetry or data collection.Needs: The prebuilt binaries need no installation. A dedicated GPU is optional, and CPU-only works more slowly. The README gives no RAM figures.Models are not included: you bring GGUF files, typically from Hugging Face.

A web interface for a runner1

A browser-based chat screen that sits on top of one of the runners above.

• Open WebUIOpen WebUI project · Web interfaceA self-hosted web interface that you open in a browser. It connects to a runner such as Ollama, or to an OpenAI-compatible server, and does not run models itself.License: Its own Open WebUI License. Code up to v0.6.5 is BSD-3-Clause. Since v0.6.6 (April 2025) a license with a branding clause applies, which its own license page says is not OSI-approved open source. It is free to use if you keep the branding or have 50 or fewer users in 30 days.Cost: The pages we opened state no price.Works on: Installed with Docker, pip (Python 3.11), uv or Kubernetes.Privacy: The FAQ says it does not collect your data or send it to external services by default. Connecting an external provider sends your prompts to that provider.Needs: No hardware requirements are stated on the pages we opened.You need a runner set up first, plus Docker or Python, so it is a second step rather than a first one.

Download models with care

1. Know what the scanner does and does not do. Hugging Face scans every file for malware, but its own page on pickle files says the scan is not 100% foolproof and that you remain responsible for checking safety.
2. Prefer safetensors over pickle. Hugging Face warns that loading a pickle file can run arbitrary code, and the Transformers docs describe safetensors files as more secure.
3. Be careful with "trust remote code". The Transformers docs say that setting turns on running code from the Hub on your machine. Use it only for repositories you trust and whose code you have read.
4. Read the license on the model card. The license field shows values such as apache-2.0 or mit, and "other" means you should read the license file. Some models are gated: you need an account and approval from the authors, and the request form shares your username and email with them by default.
5. Start with a small model. Try a size your memory handles comfortably before downloading something large.