23 Daily — JustVugg released an open-source cutting-edge Mixture of Experts (MoE) model engine named colibri, which reached GitHub Trending on November 2, 2026. Written purely in C with zero dependencies, colibri enables running enormous MoE models with hundreds of billions of parameters on existing hardware. The project offers free downloads and usage (according to the project's description).
Dynamic Expert Loading Enables Large Models on Consumer Hardware
Colibri leverages MoE models' property where only a subset of experts is used for each token, keeping a small resident portion in RAM and streaming required experts from disk on demand. This approach allows massive models to run on standard consumer PCs, requiring at least 8 GB of RAM and around 22 GB of free disk space.
The engine supports thirteen models today, including ten language models such as GLM-5.2/5.3 (also running Mistral-Large-4), several Qwen models (Qwen3.8-Flash-Next, Qwen3.6, Qwen3-Coder), DeepSeek, Inkling, OLMoE, and more. Additionally, it supports Qwen-Image-2.1 for image generation and decision models Laya, GLiNER2.5-Decide, and Clef.
Multi-platform Backend Support and User-friendly Installation
Colibri supports CPU, Vulkan GPU, and NVIDIA CUDA GPU backends, selectively compiling GPU-accelerated builds if hardware meets the criteria. Installation involves downloading a ZIP file and running a batch or shell script that detects hardware, recommends suitable models, downloads required components, and builds the engine automatically. Interrupted downloads can resume seamlessly.
An AI coding assistant script is also provided to automate the entire setup process. Users can customize model choice, backend selection, and directories via command-line options. The engine exposes functionality through both CLI and a browser dashboard, catering to various user preferences.
Editor's take: Enabling local execution of advanced MoE models expands creative AI development opportunities
Colibri's lightweight pure C implementation and dynamic expert loading from disk significantly reduce hardware barriers to access state-of-the-art large MoE models. This enables entrepreneurs with ordinary desktops or laptops to deploy powerful, flexible AI models locally. Its multi-backend support further enhances platform flexibility.
Entrepreneurs can verify colibri by installing it on their existing hardware to test model inference speed and resource usage, assessing its applicability for language, vision, or decision-making applications. The command-line tools facilitate monitoring and tuning. Its growing popularity on GitHub, with over 46,000 stars, underscores strong interest and market potential in accessible large-model AI engines.