Colibri

Colibri is an open-source inference engine that runs very large mixture-of-experts (MoE) models on the hardware you already own. With more than 46,000 stars on GitHub and an Apache-2.0 license, it lets an ordinary desktop or laptop chat with frontier-class open models, no data-center GPU required.

Overview

A mixture-of-experts model is huge on disk but small per token: only a fraction of its parameters are active for any single token. Colibri exploits that fact. The dense part of the model stays in RAM while the routed experts stream from disk through a cache that learns which experts your work actually uses. Where a weight sits changes only speed, never the result. The engine is pure C, one file per model family, with no Python at runtime and no GPU required. Thirteen engines run today, including GLM-5.2 and GLM-5.3, DeepSeek V4 and V4.1 Flash, Kimi K3, MiMo-V2.6, Qwen3.6 and Qwen3.8-Flash-Next, and OLMoE, plus Qwen-Image-2.1 for image generation and decision models such as Laya.

Key Features

• Huge models on modest machines: run a 744-billion-parameter GLM-5.2 with as little as 8 GB of RAM, streaming experts from disk.
• Every backend you have: CPU, Vulkan on any GPU, CUDA on NVIDIA cards, and Metal on Apple Silicon, with the setup picking the fastest option your machine supports.
• One-step setup: a single script inspects your machine, recommends a model that fits, downloads it with resume support, and opens the dashboard.
• Standard APIs: an OpenAI-compatible and Anthropic-compatible server, so coding assistants and editors connect with a simple base-URL change, plus an MCP server for agent workflows.
• A dashboard that shows the work: a chat UI with reasoning blocks, a live monitor, an expert map that lights up as experts fire, and per-turn profiling.
• System One decisions: ask typed questions about a document and get a probability for every allowed answer plus a confidence score, instead of free-form text.
• Beyond chat: vision models take images, documents can be attached to a conversation, and Qwen-Image-2.1 generates pictures from text.

💳 Pricing

Colibri is free and open source under the Apache-2.0 license. There is no paid tier and no account; the models run on your own machine.

Who Is It For

Colibri is for developers, researchers, and enthusiasts who want to run large open models privately, without cloud bills or API keys. It suits anyone with a decent PC who wants frontier-class models running entirely under their own control.