llama.cpp

llama.cpp is an open-source library and set of tools for running large language model inference, including vision-language models, with minimal setup. The aim is strong performance across many kinds of hardware, whether local or in the cloud. It is a plain C/C++ implementation without external dependencies, built on the ggml library, and is licensed under MIT.

The project optimizes for many platforms: Apple silicon through ARM NEON, Accelerate and Metal, x86 with AVX, AVX2, AVX512 and AMX, and several RISC-V extensions. Quantization from 1.5-bit to 8-bit integers reduces memory use and speeds up inference. GPU backends include CUDA for NVIDIA, HIP for AMD, MUSA, Vulkan and SYCL, and hybrid CPU plus GPU inference lets you partly accelerate models larger than available VRAM.

It provides a command-line interface and llama-server, which offers a REST API and a built-in web UI. You can install it from pre-built binaries, Docker or by building from source. Other local-AI tools such as Ollama build on it, which makes it relevant to developers embedding inference in their own applications and hobbyists running models on modest hardware.