colibri
SHA-256A pure C, zero-dependency engine that runs frontier MoE models locally by streaming experts from disk—no GPU required, with OpenAI-compatible APIs and a live dashboard.
Smart Download
Download Download Version
v2.0.0 · 10.5 MB
Run 700B+ parameter MoE models on a machine with as little as 8 GB of RAM.
Core Features
- Pure C, zero dependencies, one file per model family—tiny and auditable
- Streams experts from disk: a 744B model runs with only ~10 GB resident RAM
- No GPU required; optional Vulkan/CUDA acceleration when available
- One-step setup detects hardware, recommends a model, and resumes interrupted downloads
- Browser dashboard, OpenAI/Anthropic-compatible API, and MCP server included
What It Can't Do
- •Requires at least 8 GB of RAM and 22 GB of free disk for the smallest model; larger models need much more disk. Speed depends on disk read performance—an SSD is strongly recommended, and it will be slower than full in-memory inference. On WSL, never put models under /mnt/c (many times slower). First-time model downloads are large but resumable. Integrated GPUs only speed up a few models (e.g., Qwen3.6), so don't expect discrete-GPU-class performance.
Use Cases
- Chat locally with huge MoE models like GLM, DeepSeek, and Kimi on consumer hardware
- Serve a local OpenAI-compatible endpoint for other apps on your machine
- Research and teaching: watch every expert fire in real time on the dashboard
colibri is a tiny, pure C inference engine with zero dependencies that runs very large mixture-of-experts (MoE) language models on hardware you already own. Instead of loading the whole model into memory, it keeps only the dense part in RAM (about 9.9 GB for the 744B-parameter GLM-5.2) and streams the routed experts from disk on demand through a learning cache. Thirteen engines cover models such as GLM-5.2/5.3, Kimi K3, DeepSeek V4 Flash, Qwen3.6, MiMo, OLMoE, plus image and decision models. A one-step setup script detects your hardware, recommends a fitting model, builds or downloads the engine (CPU, Vulkan, or CUDA), and starts a dashboard with OpenAI- and Anthropic-compatible APIs. Compared with llama.cpp or Ollama, its single-file C engines are far smaller and easier to audit, and its disk-streaming design lets an 8–16 GB RAM machine run models that would otherwise need hundreds of gigabytes of VRAM.
Tags
Getting Started
Download installer
Click the button above to download the installer for your system
Install the software
Double-click the downloaded installer and follow the prompts
Download and unzip the ZIP from GitHub, or git clone the repository
On Windows double-click START-HERE.bat; on Linux/macOS run ./start-here.sh
Press Enter to accept the recommended model and wait for the build and download
- Download and unzip the ZIP from GitHub, or git clone the repository
- On Windows double-click START-HERE.bat; on Linux/macOS run ./start-here.sh
- Press Enter to accept the recommended model and wait for the build and download
SHA-256 checksum verified
Checksum extracted from GitHub official Release page
SHA256 Checksum
fdb8ea2fad7dcfe44d9d37bfb3e444f1299207f268ab499719f98bd772397d79This checksum is extracted from the GitHub Release page. Verify file integrity after download.
All SHA-256 checksums on this platform are extracted from the project's official GitHub Release page, without any modification. You can independently verify them on the GitHub Releases page.
Open Source Transparency
View GitHub SourceUninstall Info
Run c/coli stop first (c\coli.cmd stop on Windows), then delete the colibri source folder and the models directory (~/colibri-models by default, or the folder passed via --dir). No system-level leftovers remain.
No Extra Dependencies
Ready to use after download. No additional runtime required.
Having issues? Check the FAQ below
5 FAQs
Similar Projects
ollama
Ollama lets you download, run, and manage large language models locally. One command, multiple platforms, endless possibilities.
llama.cpp
High-performance LLM inference engine in C/C++ with minimal dependencies, supporting quantized models (1.5–8 bit) and diverse hardware (Apple Silicon, CUDA, Vulkan, etc.).
mempalace
Open-source local-first AI memory system with verbatim storage and semantic search — zero API calls, no keys, top benchmark scores.