is ExLlama free? yes — and it's the fastest way to run a local model, with one big condition

The short answer

ExLlamaV2 is a free, open-source (MIT) inference library optimized for running LLMs on consumer GPUs using the EXL2 quantization format. On a GPU it's roughly twice as fast as llama.cpp for token…

Cost: ExLlamaV2 is free and MIT-licensed. So is llama.cpp. Neither costs anything.

Speed: ExLlamaV2 is ~2x faster on a GPU with EXL2 models — measured around 14k tok/s prompt processing vs ~7.5k for llama.cpp in one test.

The catch: ExLlamaV2 is GPU-only. The whole model must fit in VRAM — no CPU offload, no running models bigger than your card.

llama.cpp wins on flexibility: GGUF format, runs on CPU, Mac, or GPU, and can offload big MoE models to system RAM.

Pick ExLlamaV2 for max GPU speed when the model fits; pick llama.cpp for flexibility, CPU/Mac support, or running models bigger than your VRAM.