BitNet 拥有一个秘密的 API 服务器。没有人告诉你。
📄 中文摘要
BitNet 是微软推出的 1-bit LLM 框架，实际上是 1.58-bit，采用三元权重，每个参数为 {-1, 0, +1}。其宣传口号是：在 0.4 GB 内存中运行 20 亿参数模型，速度比 llama.cpp 快 2-6 倍，能耗减少 82%，且无需 GPU。尽管该模型表现良好，获得了 35,000 名开发者的关注，但其生态系统却几乎不存在。目前存在 269 个未解决的问题，100 多个未合并的 PR，只有三名活跃的维护者，缺乏 Docker 镜像、pip 安装、LangChain 集成、LlamaIndex 适配器和 MCP 服务器等支持。
📄 English Summary
BitNet Has a Secret API Server. Nobody Told You.
BitNet is Microsoft's 1-bit LLM framework, technically 1.58-bit, utilizing ternary weights with parameters of {-1, 0, +1}. The framework claims to run a 2 billion parameter model in just 0.4 GB of memory, achieving speeds 2-6 times faster than llama.cpp on CPU and reducing energy consumption by 82%, all without requiring a GPU. Despite its impressive performance and garnering 35,000 stars from developers, BitNet lacks a supporting ecosystem. Currently, there are 269 open issues, over 100 unmerged pull requests, and only three active maintainers. Additionally, there are no Docker images, pip installation options, LangChain integrations, LlamaIndex adapters, or MCP servers available.