by Nomic
Free MIT-licensed desktop app that runs open-weight LLMs offline, with LocalDocs chat over your own folders. About 77K GitHub stars.
Pick GPT4All if you want a private, no-account chat app that reads your own documents on a laptop. It needs 16 GB of RAM for most models, and its latest release, v3.10.0, shipped on 2025-02-25. Newer runners such as Jan and LM Studio add faster updates and more model options, so check them too.
LocalDocs, a no-account way to chat with a folder of files using on-device embeddings and visible sources.
The licence is permissive and several local runners load the same GGUF files, so lock-in is low. What remains is name recognition from 2023 and a simple LocalDocs workflow.
Main trade-off
Release cadence has slowed, with the last release in February 2025, while newer runners add model support faster.
The desktop app is free and no paid plan is listed on the GPT4All page or the GitHub repository as of 2026-10-06. Real costs are local, mainly disk space for model files and a machine with enough memory. Nomic's separate platform for architecture and engineering firms is a different product.
GPT4All is a free, MIT-licensed desktop app from Nomic that runs open-weight language models on Windows, macOS and Ubuntu with no cloud account. Its GitHub repository had about 77,400 stars on 2026-10-06. LocalDocs lets you chat with your own folders, and an optional OpenAI-compatible server listens on port 4891 when enabled.
Offline local models
Downloads GGUF models from Hugging Face through an Explore Models page and runs them on a llama.cpp backend with no internet connection.
LocalDocs
Indexes a folder with on-device embedding models, retrieves up to 3 snippets per prompt by default and lists the source files under each answer.
OpenAI-compatible local server
Exposes an OpenAI-style API on port 4891 for other apps on your machine, switched off until you enable it in settings.
Python client
Installs with pip install gpt4all and loads a model such as Meta-Llama-3-8B-Instruct in a few lines of code.
Device and GPU control
Lets you pick Auto, Metal, CPU or GPU for inference and set how many model layers go to VRAM, 32 by default.
Opt-in data sharing
The Datalake option for sharing anonymous interactions with the community is off by default.
Most users expect these and they are not currently supported.
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| Free | Yes | Mac · Windows | -- | |
| $20/mo | Yes | Web · Mac · Windows | 4.0 | |
| $20/mo | Yes | Mac · Windows | -- | |
| Free | Yes | Mac · Windows | -- |
MIT (app); each downloaded model carries its own licence
Nothing is charged for the app itself as of 2026-10-06. What you spend is disk space for downloaded model files and a computer with enough memory to run them. Nomic also sells an unrelated business product, which you never need for GPT4All.
The limit is your hardware, not an account cap. Most models want 16 GB of memory (8 GB is enough for 3B ones), and Nomic suggests a graphics card with 8 GB of VRAM for comfortable speed. Out of the box the context length is 2,048 tokens and LocalDocs pulls 3 snippets per prompt, both adjustable.
Ollama suits developers who live in a terminal and call models from scripts. LM Studio and Jan are graphical runners with their own model search, and Jan can also connect to cloud providers. If you prefer not to run anything locally, a hosted assistant such as ChatGPT or Claude avoids the hardware requirement.
Development has slowed. The newest tagged build on GitHub is v3.10.0 (2025-02-25) and its latest push was 2025-05-27. The app still downloads and runs models, but newer runners ship updates more often, so check the repository before relying on it for a brand-new model family.
Download the installer for Windows, Windows on ARM, macOS or Ubuntu from the GPT4All page, then open Models and choose Add Model to download one. Start with a model in the 3B to 8B range if you have 16 GB of RAM or less. To chat with your own files, add a LocalDocs collection that points to a folder.
GPT4All is Nomic's free chat app for running open-weight models entirely on your own computer. The code is published under the MIT licence at github.com/nomic-ai/gpt4all, and on 2026-10-06 the repository showed about 77,400 stars and 8,300 forks, which makes it one of the most widely starred local chat apps. Its tagline is a private, local chatbot: no API calls and no GPU are required, and the vendor states that no data leaves your machine.
The app installs on Windows, Windows on ARM, macOS and Ubuntu, and a chat window lets you download models from a built-in Explore Models page. Under the hood it uses a llama.cpp backend and loads GGUF files from Hugging Face, so it can run families such as Llama from Meta and the R1 distillations from DeepSeek named in its README. The docs describe the sweet spot as models in the 3B to 13B parameter range on consumer hardware. If you want to compare it with other local runners, Ollama is built around a command line and a local API, LM Studio is another graphical runner with model search, and Jan adds cloud providers beside local models.
The feature people mention most is LocalDocs. You link a folder, the app indexes it with Nomic's on-device embedding models, and each answer can show which files it drew from under a Sources button. The default pulls up to 3 snippets per prompt. It is a private cousin of document tools such as NotebookLM, with the difference that files and chats stay on disk. An optional local server speaks an OpenAI-compatible API on port 4891, and it is switched off until you enable it. A Python client installs with pip install gpt4all.
The app is free, and neither the repository nor the GPT4All page lists a paid plan. The cost is hardware: the system requirements list 16 GB of RAM as the minimum for most models (8 GB for 3B models) and an 8 GB VRAM graphics card as the recommended spec. Expect a slower pace of change than the newer runners, because the latest GitHub release is v3.10.0 from 2025-02-25 and the last commit on the main branch was 2025-05-27. Nomic itself now markets a separate platform for architecture and engineering firms, which is a different product from this app. If you need frontier quality without the hardware, hosted assistants such as ChatGPT and Claude remain the alternative, and the models themselves are catalogued on Hugging Face.
Product facts on this page (pricing, plans, integrations, compliance) are read from Nomic's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.