OllivoChat the way you're used to
The answer appears as it's written; tables and code are formatted, with a “Copy” button. Conversations are saved and searchable. Roles and answer style — “Precise · Balanced · Creative” — instead of cryptic numbers.
early version Windows · NVIDIA graphics card
AI models at home.
No cloud, no setup.
Ollivo is a Windows app that lets anyone run open AI models — no technical skills needed. Install it, add a model, start talking. Everything runs on your computer.
No Python, Git or CUDA
Conversations stay offline
Updates itself
Works with models from HuggingFace
LM Studio
Ollama
ComfyUI
Powered by llama.cpp and whisper.cpp
Get started
Three steps to the first answer
Everything the model needs, the app downloads and sets up by itself.
Install
The wizard checks your computer and graphics driver and suggests a disk with enough space. The app installs everything else itself — no command line, no administrator rights.
Add a model
Any model you like: drag the file into the window — the app figures out what it is and whether it will run. Models you already downloaded are picked up without copying. Or take a tested one from the picks.
GGUF fileHuggingFaceLM StudioOllamaComfyUI
Talk
GPU layers, conversation memory and the other settings are tuned to your hardware. You just type — or speak into the microphone.
Features
Everything that already works
Only what the app has today. Images, video and voice-over are planned — see “What's next” below.
Documents
Attach a contract, a report or a book — the model reads it and answers from the text. If a document is longer than the model can hold, the app warns you before sending instead of silently cutting it.
Photos
Show the model a picture and ask what's in it. Phone photos are rotated the right way. If the model can't see images, Ollivo suggests a suitable one that can.
Voice
Dictate your question instead of typing it. Attach a meeting recording or a Telegram voice message — you get the text, and you can ask questions about it.
Sets itself up
How many layers to give the graphics card and how much conversation to keep in memory — the app decides. Not in use for 10 minutes — the model frees the graphics card and wakes up on the next question.
Project folder
Open a folder — the model sees its files, finds what it needs, edits them piece by piece and creates new ones. Any change by the model can be undone with one “Restore” button: copies are kept for 30 days. The model can't run programs.
Manualevery change — with your permission
Autothe model changes files by itself
Planonly a plan, touches nothing
Errors you can understand
Instead of “CUDA error 700” — what happened in plain words and a button that fixes it: “Run lighter”, “Try again”, “Install components”. Didn't help — “Report a problem”: you see the whole report before sending, and your conversations are not in it.
Honestly
Know how it will run before you download
A model that answers one word a minute isn't a “powerful model” — it's a wasted evening and ten gigabytes on your disk. So every variant in the catalog shows right away whether it will run on your computer.
- A traffic light for every modelCalculated for your graphics card and memory, not for an average computer.
- Speed in words, not “tokens”With a comparison: faster than you read, or slower.
- Compression in one phrase“Almost like the original, at half the size” instead of “Q4_K_M”.
Qwen3.5 4B
2.6 GB the usual choice: almost like the original, at half the size fits entirely in the graphics card writes about 52 words per second — faster than you readQwen3.5 9B
6.1 GB a bit better than usual and a bit bigger fits partly — will be slower writes about 10 words per second — faster than you readQwen3.5 27B
20.9 GB almost indistinguishable from the original fits, but will answer very slowly writes about 1 word per second — slower than you readThe app's real calculation for a computer with a GTX 1080 (8 GB), for English text. Your numbers will differ, and if a model won't fit at all, the light turns red with advice to take a smaller version.
Privacy
Your conversations stay with you
The model runs on your computer. Conversations, documents, photos and recordings aren't sent anywhere — they're processed right here. Ollivo collects no statistics about you. Once the model is downloaded, you can turn the internet off — the chat keeps working.
This site has no trackers or cookies either.
Ollivo goes online only to
↓
Download a model
From HuggingFace or the mirror you chose.
↓
Download engines
From GitHub and the Microsoft site; every file is checked against its checksum.
↻
Check for updates
If checking is on. Updates are signed: the app won't install a tampered one.
↑
Never leaves your computer
Send a problem report
Only when you click “Send” yourself, having seen what's in it.
✕
Conversations, files, photos and recordings
Not in reports, not “to improve the service”.
What's next
The road to version 1.0
One window for every kind of model: text, images, video and sound. The first part is ready — conversation, voice and files.
0.1
now
Conversation
Chat, documents, photos, voice, project folder, one-click report. English and Russian.
0.5
next
Images
Drawing from a description without words like “sampler” and “CFG”. Refine, extend, variations.
0.6
next
Tasks
One click: remove a background, enhance a photo, translate a document. Images right from the chat.
0.8
next
Sound and video
Reading answers aloud, talking by voice “like a call”, short videos from photos.
0.9
next
“Pro”
All parameters, logs, the ComfyUI graph, your own API — for those who need more than simple mode.
1.0
goal
For everyone
Tested on dozens of computers, installation without internet.
Download
Installs in a couple of minutes
Builds are published on GitHub. You need Windows and an NVIDIA graphics card; the more memory it has, the smarter the models it can run — the app shows exactly which ones.
1. Download the installer
Just a few megabytes. The app downloads the engine and the model itself on first launch.
SHA256 checksum — to make sure the file hasn't been tampered with:
2. If Windows warns you
Ollivo doesn't have a paid Microsoft signature yet, so a blue window may appear on first install. It's not a virus: every build has a published checksum — you can use it to make sure the file is the right one, and check it on VirusTotal.
- Click “More info”.
- A “Run anyway” button appears — click it.
FAQ
The short version
What computer do I need?
Windows and an NVIDIA graphics card. Small models run even on an old card with 4 GB of memory, or on the processor alone — slower. With 8 GB of video memory, capable everyday models already work. The app checks your computer and shows in the catalog what will run on it.
What about AMD, Intel graphics and Mac?
Not yet. First we're polishing one combination — Windows and NVIDIA — so that everything works on it without help. AMD, Intel and macOS come after version 1.0.
Where do I get models?
Anywhere you like. Ollivo works with models in the GGUF format: download any from HuggingFace or elsewhere and drag the file into the window. If you already have LM Studio, Ollama or ComfyUI, the app finds their models itself and uses them without copying — no doubled disk space.
Don't know where to start? The catalog has a small selection of tested models and a HuggingFace search.
Do I need the internet?
Only to download the app, the engine and a model. After that, chat, documents, photos and voice work offline.
How much space does it take?
The app and the engine take tens of megabytes. Models take most of the space: from 1.3 GB for small ones to 15–20 GB for big ones. Speech recognition (~470 MB) downloads only when you first click the microphone. The program folder can be moved to another disk in the settings.
How is this different from LM Studio or Ollama?
They're great, but made for people who know what “quantization” and “context” mean. Ollivo is for everyone else: plain words instead of jargon, a traffic light before you download, errors with a “fix it” button. Plus documents, photos, voice and a project folder in one window — with images, video and voice-over to follow.
Is there a Russian version?
Yes — the app and this site work in English and Russian. The app picks the language of your Windows on first launch; you can change it at any time in the settings.
How much does it cost?
Nothing for now: it's an early version, and your feedback matters more to us. If something doesn't work, click “Report a problem” in the app.
Want to support the project with money — here's how.
Support
Ollivo is made by one person
The app is free, with no ads and no tracking. If it has been useful to you, help it grow.
BoostyA monthly subscription for the project
YooMoneyOne-time, any amount, by card — convenient in Russia
Crypto
Send only the listed coin and only on the listed network — otherwise the transfer won't arrive and can't be returned. After pasting, check the first and last 4 characters of the address.
USDT TRON network (TRC20)
Minimum — 0.005 USDT.
TGEzLrBGuovgiRe9dqcm1i23ZJYXyg5LL7
TON TON network
Some exchanges call the coin GRAM.
Minimum — 0.005 TON.
UQA6f8IIN_z6hMVhKffPcLCOE91D5x7wFxHQSgnJ8wmM4Uls
Bitcoin Bitcoin network
Minimum — 0.00001 BTC.
1RnUJjdzXEqPs5t735xU9aWJAtnHGqDtq
Where the money goes
- AMD and Intel graphics cards for testing. Right now the app works only with NVIDIA.
- Server and traffic. App updates are served from our own server.
- The ollivo.ru domain. A short address for the site and updates.
- Time. The more support, the sooner images, voice and video arrive.
Helping without money
- Tell friends who wanted to try AI on their own computer about the app.
- Found a bug — click “Report a problem” in the app or write on GitHub.
- Star the repository on GitHub — it helps others find it.
Your own AI on your own computer — in a couple of minutes
Download Ollivo and ask your first question. And if something goes wrong — tell us, we'll fix it.