中文版

colibri（GitHub：https://github.com/JustVugg/colibri ，2026-07-01 创建，今晨抓取时已 46,557★）是纯 C、零依赖的 MoE 大模型推理引擎，口号是“Tiny engine, immense model”：在你已有的硬件上跑前沿 MoE 模型。

它的原理来自 MoE 的稀疏性：一个几百 B 参数的模型每次生成 token 只用其中一小部分，于是密集部分（注意力、共享专家、嵌入）常驻内存，路由专家留在磁盘、被路由器点到时再经缓存读入；GPU 有的话放最热的专家与密集层。README 列出 13 个引擎：GLM-5.2/5.3、Kimi K3、DeepSeek V4 Flash、DeepSeek V4.1 Flash、Qwen3.6、Qwen3.8-Flash-Next、Muse Spark Open 等语言模型，另有 Qwen-Image-2.1 作图与 Laya、GLiNER2.5-Decide 等决策模型；可直接读已发布的 GGUF 文件，CPU 可跑，也支持 Vulkan（任意 GPU）与 CUDA（NVIDIA 卡），另有 Apple Silicon 相关支持。最低要求是 8GB 内存（16GB 以上更好）、最小模型 22GB 磁盘空间；macOS 用 Homebrew 装 libomp、git、python 后跑 ./start-here.sh，安装程序会按本机配置推荐模型；另提供 OpenAI 与 Anthropic 兼容地址，Apache-2.0 开源。

为什么值得了解

• 本地跑大模型长期卡在“小模型才行”的印象里；colibri 把 MoE“每次只激活一小部分”的特性用到了极致，让几百 B 参数的模型在普通电脑上成为可能，速度换内存、精度与路由语义不偷偷改变
• 对已有电脑、不想租云 GPU 或按量付费的人很友好：安装程序先看机器再推荐模型，下载可断点续传，OpenAI 兼容地址意味着现有 agent 工具链不用改就能接
• 但要理性看待“几百 B”：它依赖磁盘流式读专家，快 NVMe 帮助最大；上手前先用 –list 看自己的机器到底适合哪个模型，别只看标题数字

English Version

colibri (GitHub: https://github.com/JustVugg/colibri , created 2026-07-01, 46,557 stars at this morning’s fetch) is a pure-C, zero-dependency MoE model inference engine whose motto is “Tiny engine, immense model”: run frontier MoE models on hardware you already own.

Its principle comes from MoE sparsity: a model of hundreds of billions of parameters uses only a small part of itself for each token, so the dense part (attention, shared experts, embeddings) stays in RAM while routed experts stay on disk and are read through a cache when the router asks for them; a GPU, when present, holds the hottest experts and the dense layers. The README lists 13 engines: GLM-5.2/5.3, Kimi K3, DeepSeek V4 Flash, DeepSeek V4.1 Flash, Qwen3.6, Qwen3.8-Flash-Next, Muse Spark Open and other language models, plus Qwen-Image-2.1 for images and Laya and GLiNER2.5-Decide for decisions; it can read already-published GGUF files as-is, runs on CPU, and supports Vulkan (any GPU) and CUDA (NVIDIA cards), with Apple Silicon support as well. Minimum requirements are 8 GB of RAM (16 GB or more is better) and 22 GB of disk for the smallest model; on macOS, install libomp, git and python via Homebrew and run ./start-here.sh, and the installer recommends a model based on your machine; it also exposes OpenAI- and Anthropic-compatible endpoints. Apache-2.0 licensed.

Why it’s worth knowing

• Local model serving has long been stuck at “only small models work”; colibri exploits MoE’s “only a small part activates per token” property to the full, making hundred-billion-parameter models possible on an ordinary computer — trading speed for memory, without silently changing precision or router semantics
• Friendly to anyone who already owns a computer and does not want to rent cloud GPUs or pay metered bills: the installer inspects your machine before recommending a model, downloads resume where they stopped, and the OpenAI-compatible endpoint means existing agent tooling needs no changes
• Keep the “hundreds of billions” claim in perspective: it relies on streaming experts from disk, where a fast NVMe helps most; check with –list which model your machine actually fits before diving in, rather than trusting the headline number