Local AI
What can your computer run?
Convira's local engine keeps the model on your own computer. Which models fit is decided by one thing, your memory. Tell us your computer and this page does the same arithmetic the app does.
Tell us your computer
Three ways in. Any of them ends in the same answer.
The check runs in your browser. Nothing about your computer is sent to us.
31 local models · 64,000 tokens of conversation memory, always
Every local model, by the memory it needs
Ordered by the memory each model needs to hold a 64,000 token conversation.
- One line lands here, at your computer's own memory budget.
5.2 GB the smallest local model, Qwen 3.5 4B, at 64,000 tokens
Memory
Why memory decides what you can run.
Everything a model has learned has to sit in your computer's memory the whole time it is working, next to the conversation it is holding. What fits in that share is what you can run.
The share your computer gives it
On a Mac, Convira gives the model 65 percent of your memory and keeps the rest for your operating system. On a PC with a graphics card it gives the model 65 percent of the card plus about a fifth of your system memory, and tells you when part of the model runs on the processor.
Billions tell you almost nothing
One 14 billion model on this list needs 21 GB before it will start. One 29.6 billion model needs 20 GB. The difference is how the model holds a conversation, not how big it is. That is why this page is a list of gigabytes rather than a list of billions, and why we compute the answer instead of printing a table.
The floor, and what sits below it
Every run gets at least 64,000 tokens of conversation memory, so a model only counts as fitting when there is room for the conversation beside it. When nothing on the list fits by that arithmetic, Convira recommends the cloud engine and tells you why.
That is a recommendation, not a verdict. You can still run a model on the computer in the app, and the app will let you. This page checks memory, not free space; the app checks both, so it can turn down a model this page allowed if the download will not fit.
Private Box
More than one computer holds.
A Private Box is your team's own server. It runs the same list a desktop runs, with more memory behind it.
An ordinary strong server
A machine that stays on, with no graphics card required, runs the same list a desktop runs, just with more memory: the 80 billion to 128 billion class on a 128 GB to 256 GB machine.
The frontier models
These need a server with two or more datacentre graphics cards, and their downloads run from 242 GB to 756 GB. That is the honest shape of it, and it is why this page stops at what one computer holds.
Questions
The things people ask us next.
The answers to almost every message we get about hardware.
Does a bigger number of billions always need more memory?
No, and this is the part almost everyone gets wrong. A model's parameter count tells you how big it is, not how much memory it needs to hold a conversation, and on this list a smaller model sometimes needs more memory than a much larger one. That is why the page works the answer out for your computer instead of sorting models by their name.
Do I need a graphics card?
No. Without one the model runs on the processor. A long request waits minutes rather than seconds before the first word, and Convira is built to wait rather than give up. Two cards do not add up here either: Convira loads the model on one card, so a second card does not raise what fits.
What if my computer is not on the list?
Pick your graphics and your memory instead. Any computer with the same graphics and the same memory gets the same answer. You can also enter the parts yourself, and both doors end in a real answer rather than a shrug.
My computer is not one Convira installs on. What now?
Convira installs on Apple Silicon Macs running macOS 12 Monterey or newer, on Windows 10 and 11 with an x64 processor, and on Ubuntu 22.04 or newer, Debian 12 or newer and Fedora 38 or newer on x86_64. On anything else there is no build to run, so this page gives no model answer.
What if this page recommends the cloud for my computer?
Then the cloud engine is the better choice on that computer, and you can still run a model on it. The app offers the local engine anyway, because this page works from the memory your computer is sold with and the app works from the machine in front of it. The recommendation is advice, not a lock.
How much disk do the models take?
From about 3 GB to about 76 GB. The app shows the size before it downloads. We cannot see your free space from a web page, so the app can still turn down a model this page allowed.
Does Convira measure how fast my computer is?
No. We publish whether a model fits. We do not publish tokens per second, because we have not measured your computer, and a number we have not measured is not worth printing.
Can one model need two very different amounts of memory?
Yes. A few models ship in more than one build, and the builds hold a conversation differently, so one build can need several times the memory of another for the same model. The app offers whichever build your computer can hold, and this page shows that one rather than the average of the two.
Why does a model need more memory than its download?
Because the download is only what the model has learned. In memory it holds that plus the conversation it is working on, and the memory figure counts both. The smallest model on this list downloads 3.1 GB and needs 5.2 GB of memory.
Where do these numbers come from?
The same list the app uses. The page holds each model's download size and the memory its conversation takes, and works every figure out from those two, so there is no second table to fall out of date. The sources for each chip and card are published beside the data.