Local LLMs on 8 GB RAM.
Updated Sep 7, 2026
The floor.
8 GB RAM without a GPU can run 1B–3B class models. It cannot run a comfortable 7B chat with a desktop environment and a browser also open.
Ollama is the fastest install. llama.cpp is the engine you reach for when a build flag matters. llamafile is the USB-stick path.
The shell around the model.
Open WebUI wants Docker and is happy talking to Ollama on the host. Jan is a desktop window. AnythingLLM is the document Q&A layer. GPT4All still matters on CPU-only catalogs.
ComfyUI, vLLM, and text-generation-webui are the wrong class of tool for this machine.
What this machine can hold.
8 GB RAM without a GPU can run 1B–3B class models. Larger chat windows with a desktop and browser open will thrash. Prefer the smallest documented model for the install path you choose.