The Agentic Post
Breaking
Perplexity vs ChatGPT Search vs Google AI Mode: Which AI Search Is Best in 2026?  ·  What Are Vision-Language-Action (VLA) Models? How Robots Learn to Follow Instructions  ·  How to Translate Documents With AI Accurately (and When Not To)  ·  MCP vs A2A: The Two AI Agent Protocols Explained in Plain English  ·  Is Grok Accurate for News? What Studies Show About AI Search Citations  ·  Is AI a Bubble? 7 Warning Signs Investors Are Watching in 2026  ·  
Home/Guides/How-To
How to Run a Giant AI Model on Your Own Computer With colibri

How to Run a Giant AI Model on Your Own Computer With colibri

How-To

Step-by-step guide to colibri, the free tool that runs huge open AI models like DeepSeek and Kimi K3 on a computer with 8 GB of RAM. Setup, model picks and real speeds.

You can run some of the biggest open AI models on an ordinary computer with colibri, a free open-source program that needs as little as 8 GB of RAM. The trick is that it keeps only the part of the model in use right now in memory and reads the rest from your disk. Setup is one command, but be realistic: the giant models run slowly, often a word every few seconds.

How can a huge model fit on a small computer?

Most of today’s biggest open models are “mixture-of-experts” models. Think of a hospital with hundreds of specialists: each patient only sees a few of them. These models work the same way, using only a small slice of themselves for each token they write. colibri keeps the always-needed core in RAM and pulls specialists from the disk only when called, caching the ones you use most. The project’s README notes that GLM-5.2 has 744 billion parameters but uses only about 40 billion per token.

What do I need before I start?

  • At least 8 GB of RAM, though 16 GB or more is much better
  • At least 22 GB of free disk space for the smallest model, and hundreds of GB for the giants
  • A fast NVMe SSD if you want the big models, since disk speed sets their speed
  • A graphics card is optional, colibri runs on the CPU and uses NVIDIA (CUDA) or other GPUs (Vulkan) when present

How do I install colibri step by step?

On Windows: open the colibri GitHub page, click Code, then Download ZIP, unzip it and double-click START-HERE.bat. It offers to install Python if it is missing.

On Mac: install Homebrew, then in Terminal run xcode-select --install, brew install libomp git python, git clone https://github.com/JustVugg/colibri, cd colibri and ./start-here.sh.

On Linux (Ubuntu or Debian): run sudo apt install git python3 build-essential, then clone the repository and run ./start-here.sh the same way.

The setup checks your RAM, disk and graphics card, recommends a model that fits, downloads it with resume support, and opens a chat dashboard in your browser at 127.0.0.1:8000. Next time, run the same start file and it opens straight away.

Which model should I pick?

Start small. On a normal laptop, the recommended pick is usually Qwen3.6-35B, a 23 GB download that ran at about 6 words per second on a desktop CPU in the project’s tests, and around 30 with an NVIDIA card. That feels like a usable chatbot.

The headline giants are a different experience. The README’s own measurements show DeepSeek V4 Flash at about 1 token per second on a 32 GB machine, Mistral Large 4 at about 1.3 with a good graphics card, and the 2.8-trillion-parameter Kimi K3 at about 9 seconds per token, with a 1.56 TB download. They work, but they are for patient experiments, not quick chats.

Can other apps use my local model?

Yes. colibri offers OpenAI-style and Anthropic-style web addresses on your machine, so many coding assistants and chat apps can point at it instead of a cloud service. It also has an MCP server so an AI assistant can install and manage it for you, asking before it downloads anything.

Why run AI locally at all?

Privacy and cost. Nothing you type leaves your computer, there is no monthly bill, and it keeps working offline. For a gentler first step with smaller models, see our guide to running an open-source model locally. As with any viral project, download only from the official repository.

Up Next
What Is agent-desktop? Vercel’s Screenshot-Free Computer Use for AI Agents

What Is agent-desktop? Vercel’s Screenshot-Free Computer Use for AI Agents

Computer-Use Agents

agent-desktop from Vercel Labs lets AI agents use Mac, Windows and Linux apps by reading the accessibility tree instead of screenshots. How it works and is it safe.

agent-desktop is a free, open-source tool from Vercel Labs that lets AI agents see and use desktop apps on macOS, Windows and Linux without taking screenshots. Instead of looking at pixels, it reads each app’s accessibility tree and gives the agent a clean list of buttons and fields it can click or type into. With about 19,700 GitHub stars, it was one of the most viral AI projects of early November 2026.

What problem does agent-desktop solve?

Most computer-use agents work like a person squinting at a screen. They take a screenshot, ask an AI model where the button is, then click at those coordinates. That is slow, expensive and fragile: a moved window or a dark mode switch can break it. agent-desktop skips the picture entirely.

How does it work, in plain English?

Every operating system keeps a hidden map of what is on screen so that screen readers can describe it to blind users. That map is called the accessibility tree, and it lists every window, button, menu and text box with its name. agent-desktop reads that map and returns it as structured data, giving each element a short ID called a ref.

  • The agent asks for a snapshot of an app, for example Finder or Slack
  • It picks the ref for the element it wants, such as a Save button
  • It tells agent-desktop to click or type on that ref
  • It takes a new snapshot to check the result, and repeats

The AI “brain” stays in your own agent. agent-desktop has no model inside, it is just the hands and eyes. It uses Apple’s AX API on macOS, UI Automation on Windows and AT-SPI on Linux, per its GitHub README.

What can it do?

It ships around 56 working commands: clicking, typing, scrolling, dragging, opening and closing apps, moving windows, reading notifications and using the clipboard. By default it works “headless,” acting through the accessibility system without moving your mouse or stealing focus. It installs with npm or Homebrew and also runs as an MCP server, so assistants like Claude Code or Cursor can call it as a set of tools.

How is it different from screenshot-based agents?

Speed and reliability. Text is far cheaper for a model to read than an image, and a ref points to an exact element, so there is no guessing at pixel positions. The trade-off is coverage: apps with poor accessibility support, some games, and custom-drawn interfaces expose little or nothing in the tree, so a screenshot approach can still see things agent-desktop cannot. Our computer-use agent comparison shows how the main approaches stack up.

Is it safe to let an agent control my desktop?

Treat it with care. The project itself warns that snapshots, clipboard reads and screenshots can contain private content, so keep them out of public logs. An agent that can click anything can also click the wrong thing, or be tricked by text hidden in an email or web page. Read our guide to computer-use security risks, grant accessibility permission only to tools you trust, and keep a person in the loop for anything that sends, buys or deletes.

Who should try it?

Developers building agents that must work in desktop software with no API, such as older business apps, design tools or chat clients. For everyday users it is still a building block, not a finished assistant: you need an agent to drive it.