Skip to content
Ghivo
◈

AI Chat

Text and HumanizerFree · Private

Chat with a real Llama 3.2 language model that runs entirely on your own device via WebGPU. No remote AI service — and web search, when you turn it on, stays limited to just that one query.

Downloaded once to this browser, then cached — every message after that runs entirely on-device, with nothing ever sent to a server.

A real language model, running on your device

Every other tool in Ghivo uses a small, specialized model for one narrow job. This one is different: it's a genuine general-purpose chat model (Llama 3.2) running fully in your browser via WebGPU — your own graphics hardware, not a server. Ask it anything; nothing you type is sent anywhere, unless you turn on web search for a message — see below.

How to use it

  1. Pick a model size — 1B for a faster, lighter download, or 3B for noticeably better answers.
  2. Wait for the one-time download to finish — it's cached in your browser after that.
  3. Type a message and press Enter (or click Send). Responses stream in as they're generated.
  4. Click New chat to clear the conversation, or Switch model to pick a different size.
  5. Use Copy, PDF, or Word next to the message box to export the latest reply — built right here in your browser, nothing uploaded.

Attaching a PDF or photo

Click the paperclip to attach a PDF or an image (PNG, JPG, WebP). A PDF's text layer is read directly; a photo is read with on-device OCR. Either way, the extracted text is added to the conversation so you can ask about it — nothing about the file is uploaded anywhere.

One honest limit: there's no vision model behind this, so a photo is only read for any text it contains (like a scanned document or a receipt) — the model can't describe what's actually in the picture.

Searching the web

Click the globe icon to turn on web search for your next message. This is the one deliberate exception to "nothing leaves your browser": your search query (just the query — not your conversation, not the file you attached) is sent to Ghivo's own search endpoint to fetch results. The results come back as plain links and snippets, which the model reads to answer your question and cites by number. It's off by default, and stays off for every message until you turn it on again.

Why WebGPU?

Language models this capable need real parallel compute to run at a usable speed — that's what a GPU is for. WebGPU is the browser API that lets a web page use it directly, with no native app or install. It's available in recent desktop Chrome and Edge; support in other browsers is still rolling out.

Privacy

The model downloads once from a public model registry and runs entirely in this tab from then on. Your messages, and the model's replies, stay in this browser's memory for the length of the conversation — they're never transmitted, logged, or stored by Ghivo. The only exception is web search: when you turn it on, that message's search query is sent to Ghivo's own search endpoint — see above.

Frequently Asked Questions

Is this really running in my browser, not a server?
Yes, for chatting itself — the model (Llama 3.2) downloads once to your browser and runs entirely on your own graphics hardware via WebGPU. The one exception is web search: when you turn it on for a message, that message's search query is sent to Ghivo's own search endpoint to fetch results. It's off by default.
How does web search work, and what does it send?
Click the globe icon to turn it on for your next message. Only that message's search query is sent — not your conversation history, not any attached file. The results come back as links and snippets for the model to read and cite.
Why does it need WebGPU?
Running a real language model fast enough to be usable needs your GPU. WebGPU is how a browser tab can access it. It currently works in recent desktop Chrome and Edge — support elsewhere is still rolling out.
Why is the download so big?
Unlike the small specialized models used elsewhere on Ghivo, this is a general-purpose language model — a few hundred megabytes to a couple of gigabytes depending on which size you pick. It downloads once and is cached for every chat after that.
Why did it ask me to download the model again?
Browsers can clear a site's cached storage during normal cleanup if the site hasn't earned "durable" storage status — for a cache this size, that can mean losing it after as little as a day. Ghivo now requests persistent storage the moment you pick a model, which should stop this from happening going forward; if your browser still clears it occasionally, that's a browser-level storage decision, not something tracked or caused by Ghivo.
Does conversation history get uploaded or saved anywhere?
No. The conversation lives only in this browser tab's memory for on-device inference — it's not stored, logged, or sent to Ghivo or anyone else. Turning on web search only ever sends that one message's search query, never the rest of the conversation.
Can I attach a PDF or a photo?
Yes — click the paperclip. A PDF's text is read from its text layer; a photo is read with on-device OCR. The extracted text is added to the chat so you can ask about it. There's no vision model behind this though, so a photo is only read for text it contains, not visually described.

Related Tools

◈
AI Text Humanizer

Humanize AI-generated text instantly in your browser. Includes on-device AI detection. Models run locally — your text is never sent to any server. Three rewrite modes: Light, Medium, Aggressive.

◈
Grammar Fixer

Fix grammar, spelling, and punctuation mistakes instantly in your browser. A local AI model runs on-device — your text is never sent to any server.