Run This Ai
EN DE

Colibri

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model.

★ 25,364 GitHub Apache-2.0 llminferencemoeself-hostedcopenai-compatible LLM & Chat

Overview

Colibri is a pure-C inference engine that runs frontier Mixture-of-Experts models such as GLM-5.2 (744B), Kimi K3 (2.8T), Inkling (975B) and DeepSeek V4 Flash (284B) on consumer and heterogeneous hardware. It treats VRAM, RAM and storage as a single multitier memory hierarchy, streaming experts from disk on demand, so a 744B model can run with as little as ~10 GB of resident RAM. It ships a chat CLI (coli chat), an OpenAI-compatible server (coli serve) and a live web dashboard (coli web) with token metrics, tier bars and a real-time expert cortex view. Apache-2.0 licensed, one C file per model family, zero engine dependencies.

Requirements

Min vCPU
2
Min RAM
8192 MB
Min Disk
10 GB
Rec vCPU
8
Rec RAM
32768 MB
Rec Disk
20 GB

Related tools

NextChat

Cross-platform ChatGPT web UI with multi-model support for Ollama, Claude, Gemini, and more

Lobe Chat

Extensible, open-source ChatGPT alternative with plugins, knowledge base, and multi-LLM support

Open WebUI

User-friendly WebUI for LLMs (Ollama, OpenAI API)

text-generation-webui

Run local LLMs with a powerful web interface — text, vision, tool-calling, and OpenAI-compatible API

Streamlit

Build and share data apps in pure Python — fast

Gradio

Build and share delightful machine learning apps in Python

Guides & articles

How to Deploy Colibri: Chat with a 744B MoE Model on Your Own Machine

Step-by-step tutorial: build Colibri from source, chat with a 744B GLM-5.2 model from your terminal, expose it via an OpenAI-compatible API, and deploy it with Docker Compose.

Aug 18, 2026

Colibri: Run Frontier 2.8T MoE Models on Hardware You Already Own

Colibri is a pure-C inference engine that streams frontier MoE models — GLM-5.2 (744B), Kimi K3 (2.8T) and more — from disk onto the hardware you already own. Here's how it works and why the memory hierarchy changes everything.

Aug 18, 2026