llama.cpp — Inference & Compute Platform

llama.cpp is the foundational C/C++ inference engine that redefined what's possible for running large language models outside of multi-billion-dollar data centers. With 107,000+ GitHub stars, it's the backbone of nearly every local-LLM tool — Ollama, LM Studio, GPT4All, Open WebUI, and countless others build on llama.cpp's runtime.

Its core innovations are the GGUF model format (a holistic single-file package containing weights, tokenizer config, and architecture metadata) and a comprehensive quantization stack: 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization with K-quants and IQ-quants. For coding and reasoning models, Q4_K_M or Q5_K_M is the practical sweet spot.

Hardware support is extensive: Apple Silicon (ARM NEON, Accelerate, Metal — first-class support), x86 (AVX, AVX2, AVX512, AMX), NVIDIA GPUs (custom CUDA kernels), AMD GPUs (HIP), and Moore Threads (MUSA). The project is fully open-source under MIT, maintained by ggml-org/Georgi Gerganov, and is the standard tool for local LLM inference in 2026.