🔥 Hot Repo: Native GGUF, Zero Python — 1T Models Run Locally

One-liner — Colibri is a pure C inference engine that streams mixture-of-experts model experts from disk so frontier models from 35B to 1 trillion parameters run on hardware you already own — no Python, no cloud, no dependencies. - Repo: JustVugg/colibrihttps://github.com/JustVugg/colibri - Stars: ⭐ 47,000 trending today following v2.1.0 - Language: C - License: Apache-2.0 --- What It Does Colibri is a self-contained C inference engine that runs large MoE models by loading only the expert layers each token actually routes to, keeping memory usage far below the full model weight. A model with 744 billion or more parameters that would normally require a cluster runs on a consumer machine because only a fraction of the total weights are ever in memory simultaneously. No Python runtime, no package manager, no cloud account — it unpacks, builds with make, and runs. Why It's Blowing Up v2.1.0 landed on October 31, and it crossed two lines that change Colibri from a curiosity to a credible inference platform. The first is native GGUF support. Before v2.1.0, you had to use Colibri's own model format, which meant converting files or downloading from the project's own Hugging Face organization. Now coli import model.gguf accepts any GGUF without conversion or re-quantization. That unlocks every quantized model on Hugging Face — tens of thousands of files — for use in Colibri with one command. The second is cross-platform GPU acceleration. v2.1.0 ships a Linux and Windows Vulkan backend where models larger than VRAM overflow experts into a RAM pool, a Windows CUDA archive for NVIDIA cards compute 8.0+, and an Apple Silicon macOS build with Metal. For the first time, users on NVIDIA, AMD, and Apple Silicon all have accelerated paths without compiling from source. New model support adds Mistral-Large-4 — 1 trillion total parameters, 52B active via MoE — plus Muse Spark Open 816B, DeepSeek V4.1 Flash GGUFs, and Qwen3.8-Flash-Next at 262,144-token context. The KV cache