Localynx AI
Free AI chatbot that runs in your browser with WebGPU. No account, API key or AI server: pick a model, download it once and keep chats on your device.
Still loading. If the tool doesn’t appear, reload this page.
About this tool
Localynx AI
Chat with an AI language model that runs entirely in your browser with WebLLM and WebGPU. Pick a model, stream answers, keep your chats on this device and use it offline.
Localynx AI is a ChatGPT-style assistant whose language model runs on your own computer, inside this browser tab. It uses WebLLM and WebGPU to run open models such as SmolLM2, Qwen and Phi on your graphics card, so your questions and answers are not sent to an AI server and no API key, subscription or account is needed. Choose a model, download it once (from about 200 MB for the small ones to about 2 GB for the large ones) and it is kept in your browser, so later chats start in seconds and keep working with no internet connection. Answers stream in as they are written, with Markdown and highlighted code blocks, and you can stop or regenerate a reply, ask follow-up questions that use the whole conversation, rename or delete chats, and export one as Markdown. Your conversation history is saved only in this browser. Because the models are small and run locally, answers can be shorter or less accurate than the big cloud assistants, and a recent computer with WebGPU (Chrome or Edge 113 or newer) is required. The model you choose downloads from huggingface.co, and your chats are not sent there. The model's compiled WebGPU code downloads from raw.githubusercontent.com.
How to use Localynx AI
- 1
Open the tool in Chrome or Edge on a computer with a graphics card. If WebGPU is not available the page tells you why and what to change.
- 2
Press Download on the suggested model (Qwen2.5 0.5B, about 290 MB) or open the model list to pick another. The progress bar shows the download, then the model loads onto your GPU.
- 3
Type a message and press Enter. The answer streams in; use Stop to cut it short, or the refresh icon to regenerate it. Ask follow-up questions in the same chat.
- 4
Use New chat for a fresh topic, and rename, delete or export chats from the sidebar. Settings change the system prompt, creativity and maximum answer length.
Why use Localynx AI
Chats stay on your device
The model runs on your GPU in a Web Worker. There is no AI server, API key or tracking of what you type, and chats live only in your browser's IndexedDB.
Works offline after one download
Model files are saved in the browser's cache storage. Once a model is downloaded, loading it and chatting need no network; only opening the page itself needs the site.
Honest about size and hardware
Every model shows its download size, the memory it needs and its licence, and the tool warns when a model is probably too large for your device or phone.
Free and open
Only free, open-source software and openly licensed models (Apache-2.0 and MIT) are used, with no usage limits, credits or subscriptions.
Common use cases
- You want to ask questions or draft text about something private, such as work notes, personal messages or code you are not allowed to paste into an online service.
- You are on a plane, a train or a patchy connection and want an assistant that still answers after the model has been downloaded once.
- You are comparing small open language models side by side, or learning how in-browser AI with WebGPU performs on your own hardware.
Frequently asked questions
Is my conversation sent to any server?
No. The model runs on your own graphics card inside the page and there is no AI backend. Your chats are saved only in this browser. The only network use is the one-time download of model files from Hugging Face and the MLC project.
What do I need to run it?
A recent desktop browser with WebGPU (Chrome or Edge 113 or newer, or Chrome 121+ on Android) and a computer with a working graphics adapter. The small models need under 1 GB of graphics memory; the largest ones want 4 GB or more.
How big are the downloads?
From 207 MB for SmolLM2 360M and 290 MB for Qwen2.5 0.5B up to about 2.2 GB for Phi-3.5 mini and Qwen3 4B. A model is downloaded once and then stays in your browser until you remove it from the model list.
Does it work offline?
Yes, after a model has been downloaded. Open the page while you are online, then load the model: it comes from your browser's cache and chatting needs no connection. Opening the page itself still needs the site to load, so keep it open or load it once before going offline.
Why are the answers sometimes wrong or short?
These are small models (0.3 to 4 billion parameters) running on a laptop, much smaller than cloud assistants. Pick a larger model for better answers, keep questions focused, and double-check anything important.
Which models can I use and are they free to use?
SmolLM2 and Qwen models are Apache-2.0 and Phi-3.5 mini is MIT, all free to use. Llama and Gemma are left out because their licences add restrictions.
Why does the model say it forgot earlier messages?
Each model remembers about 4,096 tokens (roughly 3,000 words) of the conversation including its own answer. Older messages are left out automatically in very long chats, and the tool tells you when that happens; start a new chat for a new topic.
Why did loading fail or the GPU reset?
Usually the model was too large for the available memory or the download was interrupted. Close other tabs, choose a smaller model, and try again; files that finished downloading are kept, so a retry resumes where it stopped.
Where are my chats stored and how do I delete them?
In your browser's IndexedDB on this device. Delete one chat from the sidebar, or use Settings, Delete all chats. Clearing the site's data in your browser also removes chats and downloaded models.