llama.cpp 免费吗？价格、免费额度与海外可用性

GGUF 与 k-quant / i-quant 量化的事实标准实现，纯 C/C++ 编写，不依赖 PyTorch。CPU、CUDA、Metal、Vulkan、ROCm、SYCL 全后端覆盖，同一份 GGUF 文件在 Windows、macOS、Linux、甚至手机上都能跑。自带 llama-server、llama-quantize、llama-bench 等工具，量化、压测、起服务全在命令行里解决。你在用的 Ollama、LM Studio、KoboldCpp，本质都是给它套了层壳。

本地推理世界的地基。想认真做量化，绕不开它：官方 README 里那张各量化档位的体积与吞吐实测表，是整个社区选档位的依据。缺点是纯命令行，文档靠 README 和 PR，新手第一天会有点懵。建议路径是先用 Ollama 跑通，需要精细控制量化参数时再回来啃 llama-quantize。

The C/C++ inference engine where GGUF and k-quant / i-quant quantization originated. Runs on CPU, CUDA, Metal, Vulkan and ROCm, from an 8B model on a laptop to 405B on a server, and it underpins most local tools — Ollama, LM Studio, KoboldCpp and others are largely wrappers around it. Ships llama-quantize, llama-bench and llama-server.

The foundation of the local inference world, and worth learning if you care about quantization.

Where it winsThe quantization toolchain is the reference — the published size and throughput tables for each quant level are what the whole community uses to choose. Backend coverage is unmatched, so the same GGUF runs everywhere.

Where to be carefulIt is command-line only, and documentation lives in the README and pull requests rather than a guide. The first day is confusing.

VerdictStart with Ollama to get running, then come back here when you need real control over quantization parameters. If you are choosing quant levels for production, you need to understand this tool.