Daimax Write

Private autocomplete in any text box, written by a small language model on your own GPU. Nothing you type leaves your computer.

Daimax Write suggests the rest of what you are writing, in grey, right where you type: emails, comments, forms, documents, any text box on any site. The suggestions come from a small open language model (Liquid AI's LFM2) that runs on your own computer's graphics chip through WebGPU. Nothing you type is sent anywhere: there is no account, no server that sees your text, no analytics and no ads. The one thing that goes over the network is the model download from Daimax's model host, a Cloudflare service; it carries nothing you type, and like any download it sees your IP address. How it works - Start typing. A fast model suggests the next few words as you go. - Pause, and a larger model drafts the rest of the paragraph. - Press Tab to take the suggestion, or keep typing to ignore it. Escape dismisses it. - It reads the page title, the field's label and a little of the page (the message you are replying to) so the suggestion fits, and never reads password fields. What it keeps stays on your device - It keeps a local record of what you write and which suggestions you keep. Neither the paragraphs it keeps nor the facts you confirm are given to the model. When a fact comes up often in them (your company, a client's name), the side panel offers it; once you confirm it, a suggestion that uses it is not stopped to ask about it, as long as learning is on in the side panel. - Small models like to make details up, so it watches its drafts for the usual ones: a person's full name, a day or date, a time, an amount like $4,800 or 30 cm, an email address, phone number, link or street address, and blanks like [your name]. When one is about to go in that nothing on the page or in your confirmed facts gives, it often stops the suggestion right there and asks, like [which day?]. Answer in the side panel and the draft is redone with your answer, which is used for that message only and not saved. Rather not say? Let AI make it up. - The side panel lists what it has kept, per site, with the full details and a delete button on every entry. You can switch it off on any site (it then does nothing there), forget a site, or forget everything. What you need - Chrome with WebGPU, on a computer whose graphics chip supports it (most Macs with Apple silicon, and most recent Windows PCs). - The first time it runs it downloads the models: about 420 MB for the fast model, and about 1.4 GB for the larger one once the fast one works. After that it works offline. Open models, credited: LFM2 by Liquid AI (LFM Open License v1.0), quantized and compiled for WebGPU by Daimax. Runs on WebLLM (Apache 2.0).