llama.cpp is the layer most local model tooling is built on. It is a C and C++ inference engine with no Python runtime, aggressive quantisation support, and hand-tuned kernels for a remarkably wide range of hardware.
It rewards people who want control. Context length, thread counts, GPU layer splits, and cache types are all exposed, and the bundled server speaks an OpenAI-compatible API for anything that needs one.
Decisive facts
- Quantisation from 8-bit down to sub-2-bit formats
- Apple Silicon, CUDA, ROCm, Vulkan, and pure CPU backends
- Bundled OpenAI-compatible server binary
- Grammar-constrained sampling for structured output
- No Python runtime required at inference time