Colibri
JustVugg/colibriReviewed by AI
Checked by RepoPicks Radar on
What is Colibri?
colibri chats with big open AI models on the machine you already own. Only the part a model needs for each word stays in memory; the rest is read from the disk as it is asked for, so a 744-billion-parameter model can start on a computer with 24 GB of RAM, slowly. It comes with a chat page in your browser, a terminal chat and a local server that other apps and coding tools can talk to. You need at least 8 GB of RAM, 22 GB of free disk for the smallest model and a download; a graphics card is optional.
Colibri is free and open source (Apache-2.0) and can replace ChatGPT and OpenRouter.
How easy is it to install Colibri?
B · installer plus one simple step · Windows
- Windows B
- macOS C
- Linux C
Windows is the easiest: you download the ZIP, double-click START-HERE.bat and, in the README's words, "If Python is missing, it offers to install it for you", then answer one question and it downloads the model and opens the dashboard in your browser. On Mac and Linux the same thing takes terminal commands and installing tools first ("sudo apt install git python3 build-essential", or Xcode tools plus Homebrew packages), so those are a C.
Colibri on GitHub
- Stars
- 49,729on Nov 10, 2026
- Downloads
- 113,90717,627 this week
- Growth
- +2,490stars in 7 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python
- To build from source
- Python, Graphics card
- Disk space
- 25 MB–74 MB · from the installer
- Accounts or keys
- None
- Cost
- Free · Free and open source (Apache-2.0); the README has a "Buy me a coffee" donation button.
- License
- Apache-2.0 (Permissive)
- Can replace
- ChatGPT Plus, OpenRouter
- Last change on GitHub
- Oct 31, 2026
Colibri risk signals (3)
- Attentionvery large downloads
The program files are small, but the models are not: the README asks for "22 GB free on the disk for the smallest model" and its tables list downloads of 23 GB (Qwen3.6-35B), 429 GB (GLM-5.2) and 1.56 TB (Kimi K3). Plan for the disk space and the download time.
AI review
- Infomodel licences restrict commercial use
The engine is Apache-2.0, but the models have their own terms: the README marks "Qwen-Image-2.1: text to picture, non-commercial licence" and "Muse Spark Open (816B, 42B active): text and tools, under Meta's licence agreement". Check each model's licence before using it at work.
AI review
- Infospeed can be very slow
The project's own measurements show how far apart the models are: "0.05-0.1 tok/s cold on a 25 GB laptop" for GLM-5.2 and "about 9.4 s per token" for Kimi K3, against 6 to 30 tokens a second for the small Qwen3.6-35B. On a normal computer the biggest models answer a word at a time.
AI review
Evidence is shown in its original language.
What Radar could not check about Colibri
- Whether the speeds in the README's tables hold on your own computer and disk
- Whether the Windows script really installs Python and sets up the engine without a hitch
- How much disk space and memory you end up using for the model you pick, plus its expert cache
- Whether the model downloads work without a Hugging Face account
- Whether the answers match the quality of the same model run on a cloud service