Appearance

Hierarchical inference engine

Your machine becomes something more.

Two hands made of individual points reach towards a small dotted computer, with the Colibri wordmark on its screen.

COLIBRI

Colibri is an open-source hierarchical inference engine that adapts to your machine to run LLMs, coordinating CPU, GPU, memory and storage.

Get started See the dashboard

01 / The Colibri dashboard

Chat with your models.
See the engine at work.

The dashboard brings local chat, expert routing, activity and structured decisions into one workspace.

Install Colibri ↓
SCROLL TO EXPLORE ↓

Your model. Your conversation. On your machine.

INTERACTIVE PREVIEW

Illustrative activity. Live metrics appear in your own dashboard.

02 / Supported models

Choose your model.
Run it locally.

Browse the models Colibri supports, their capabilities and the commands to prepare them. Every language model also runs from its published GGUF files, read as they are, unsloth's Dynamic quants included: coli import model.gguf.

21 entriesSELECT TO EXPLORE
Qwen11 / 21

Qwen3.8-Flash-Next

176B
6B activeMixture of experts

Engine documentation ↗
Prepare this model

Preparation commands vary by checkpoint. See the engine documentation for requirements.

21 catalogue entries · 13 engines. GGUFs run as published, unsloth's included.
Model weights keep their own licences.

Full project catalogue ↗

03 / How the engine works

One hierarchy.
Your hardware.

Colibri treats GPU memory, system RAM and storage as one hierarchy. It adapts where model weights live and how they are executed to the resources available on your machine.

Weights can stay in memory when capacity allows, or stream from storage when needed. CPU and GPU execution work with the hardware and backends available.

Inside the engine ↗
QWEN3.8-FLASH-NEXTONE EXAMPLE: MoE ROUTING
Selected for this tokenThe full expert set
176Btotal parameters
6Bactive parameters

One example within the engine: a MoE model selects a subset of experts for each token. This illustrates routing, not a fixed placement in storage or memory.

One engine. Different machines.
CPU · CUDA · Vulkan · Metal
Windows / Linux / macOS

04 / System One

Structured decisions,
with probabilities.

Give the model a question and the allowed answers. Get a probability for each choice.

Route a request, pick an action or score a signal. System One supports named choices, yes/no questions and defined scales.

Laya · GLiNER2.5-Decide · Clef
Meet System One ↗
THE QUESTION

What should happen to this pull request?

8 changed files. 340 lines added.
No tests included.

97.4%
TOP CHOICE

Request changes

Example outputs with illustrative probabilities. No inference request is sent.

05 / APIs, MCP and CLI

Your agents.
Your models.

Bring Hermes, OpenClaw, Pi, Codex and Claude Code to the models you run with Colibri. Keep your workflow and choose your inference backend.

OpenAI-compatibleAnthropic-compatibleMCPCLI
Build with Colibri ↗
~/colibri

# A conversation, from your terminal.

coli chat

Open the CLI and talk to your local model.

ON YOUR MACHINE

YOUR TOOLS, POWERED BY COLIBRI

Your agent. Your local model.

Choose a custom OpenAI-compatible provider in Hermes, point it to your Colibri server and select the model you are serving.

HermesYour model
Connection
OpenAI-compatible
Base URL
http://127.0.0.1:8000/v1
Model ID
Use the ID served by Colibri

Start Colibri with a prepared model. Agent workflows require a model with tool support; capabilities depend on the model and client configuration.

Connection guide ↗

And your other compatible tools: Aider, Continue, Cline and clients with a custom model endpoint.

01

Your weights.
On your disk.

02

Your inference.
On your hardware.

03

Your engine.
Open source.

06 / Install Colibri

Install Colibri.
Run your first model.

The launcher checks your hardware, guides model setup and opens the dashboard. Currently shipping v2.1.0+nui.1.

A DOUBLE CLICK TO BEGIN

Download. Unzip. Start.

  1. Open the project and choose Code → Download ZIP.
  2. Unzip the folder and run START-HERE.bat.
  3. Follow the setup for your model and machine.
Get Colibri ↗

RAM, disk space and supported backends depend on your model. Read the quickstart ↗

OR ASK YOUR CODING ASSISTANT

Let your assistant handle the first steps.

Set up colibri on this machine following docs/AI_SETUP.md from https://github.com/JustVugg/colibri
EXPLORE COLIBRI