colibri: Tiny Engines, Immense Models
Run frontier MoE models on hardware you already own
A mixture-of-experts model of hundreds of billions of parameters
Pure C, zero dependencies, experts streamed from disk.
13 engines ready — from 35B to 2.8T parameters
- 🐦 GLM-5.2 / 5.3 (744B MoE)
- 🤖 DeepSeek V4 Flash & V4.1 (284B / 552B)
- 🔥 Qwen3.8-Flash-Next (125B)
- 🎨 Qwen-Image-2.1 (text to picture)
- 🧩 Decision models: Laya, GLiNER2.5, Clef
No GPU required — CPU runs everything, GPU accelerates
- 🖥️ CPU: every engine runs with nothing else installed
- 🎮 Vulkan: any GPU with 1.2 driver
- 🚀 CUDA: NVIDIA cards with toolkit support
- 📦 Multi-drive: read experts from NVMe, two drives +37.5%
One step to get started
... ./start-here.sh — asks for your machine specs, recommends a model, builds or downloads the engine, and opens the dashboard.
... ./start-here.sh --yes — no questions, go straight to the recommendation.
System One: make models answer choices, not paragraphs
... POST /v1/systemone — gives the model a document and only allowed options, returning probabilities and confidence.
FAQ
Q: Can I run large models on a laptop?
A: Yes — on a 25 GB laptop, every expert streams from disk and runs correctly, though slower.
Q: Do I need a GPU?
A: No — every engine runs on CPU; a GPU is optional but faster for the hot experts.
Q: How do I connect colibri to other apps?
A: Yes — use the OpenAI-compatible API (base URL http://127.0.0.1:8000/v1) or the Anthropic-compatible endpoint; existing clients work unmodified.
🎉 Get started now and watch your own hardware run frontier models!
... https://github.com/JustVugg/colibri
