LM Studio vs Ollama: headless server setup on RunPod

LM Studio can run without its desktop interface. This tutorial demonstrates a headless Linux installation on RunPod, then connects a Python client to the model server. The opening comparison describes Ollama as a tool the speaker normally uses through the CLI and LM Studio as a visual model browser and chat app.

The setup uses a RunPod A40 instance, which the speaker describes as having 40 GB of VRAM. Through a JupyterLab terminal, he updates the system, runs the LM Studio installer, adds the lms CLI to the shell path, and starts the daemon. He then downloads Qwen 3.6 27B MTP in Q4 format; the download prompt reports 17.96 GB.

The walkthrough covers listing downloaded and loaded models, loading a model into memory, and testing an interactive chat. Starting the server exposes port 1234, and a request to /v1/models checks the model list. The supplied Python example uses an OpenAI-compatible API with the RunPod endpoint as its base URL.

This is a remote, self-hosted deployment rather than an offline desktop session. The speaker says his computer has an 8 GB GPU and 16 GB of RAM and cannot run the demonstrated model. He presents rented GPU hosting as a way to access it from that computer. He also recommends vLLM for handling multiple inference requests, while describing Ollama and LM Studio as options for local projects; the tutorial does not benchmark that comparison.