LM Studio: Run and Serve Local LLMs With a GUI and API Server

Why LM Studio

Most local LLM tooling assumes you are comfortable in a terminal. LM Studio removes that barrier: it is a desktop application (Mac, Windows, Linux) that handles model discovery, download, hardware tuning, and serving through a graphical interface. It is the fastest path for non-terminal users to run a model locally - and it is equally useful for developers who want a visual config layer on top of llama.cpp.

Installing LM Studio

Download the installer from lmstudio.ai. The app bundles a pre-compiled llama.cpp backend, so there is nothing else to install. On first launch, LM Studio detects your GPU (NVIDIA/AMD/Apple Silicon) and configures defaults accordingly.

Downloading Models

Use the built-in search to browse HuggingFace Hub directly. LM Studio surfaces GGUF-format models from trusted publishers like TheBloke and Bartowski. Select a quantization level - Q4_K_M is a good default - and click Download. Models are stored in ~/LM Studio/.

The supported formats page lists GGUF, MLX (Apple Silicon native), and GPTQ.

GPU Layer Offloading

In the model settings panel, the GPU Layers slider controls how many transformer layers are offloaded to VRAM. More layers = faster inference. Set it to the maximum your VRAM can hold - LM Studio shows a live VRAM meter. A rule of thumb:

Chat Playground

The Chat tab gives you a full conversation interface with message history, system prompt editor, and generation parameters (temperature, top-p, repeat penalty). You can save and load presets - useful for comparing prompting strategies across models.

Local API Server

Enable the server from the Local Server tab. LM Studio starts an OpenAI-compatible HTTP server on localhost:1234. Drop-in replacement in any app:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")
response = client.chat.completions.create(
    model="local-model",
    messages=[{"role": "user", "content": "What is GGUF?"}],
)
print(response.choices[0].message.content)

The model name "local-model" is a placeholder - LM Studio routes all /v1/chat/completions requests to whichever model is loaded.

Hardware Requirements by Model Size

For most laptops, a 7B or 8B model at Q4_K_M runs at 15 - 40 tokens/sec - fast enough for interactive use. Apple M-series chips with unified memory outperform discrete NVIDIA cards of equivalent memory size because there is no PCIe bandwidth bottleneck.

LM Studio is updated frequently - check the changelog for new features like multi-model serving and MLX acceleration on Apple Silicon.