Skip to content

LM Studio

What it is

LM Studio is a desktop application for discovering, downloading, running, and chatting with local models.

What problem it solves

It lowers the barrier to local LLM experimentation by packaging model discovery, downloads, chat, and an OpenAI-compatible local server into one desktop workflow.

Where it fits in the stack

AI & Knowledge / Local Model Workbench. It is a practical bridge between end-user experimentation and local inference.

Typical use cases

  • Testing local models without a CLI-heavy setup
  • Running a local OpenAI-compatible endpoint for development
  • Comparing small and medium models on a laptop or workstation

Strengths

  • Easy local-model onboarding
  • Friendly UI for experimentation
  • Useful stepping stone before deeper infrastructure choices

Limitations

  • Less flexible than lower-level inference stacks for production
  • Desktop-first workflow is not ideal for multi-user deployment

Apple Silicon / Metal backend

LM Studio v0.3.0+ ships with a native Metal inference backend for Apple Silicon, using llama.cpp under the hood. All 48 GB of the M5's unified memory is addressable by Metal — there is no separate VRAM pool. This makes the M5 MacBook a better local LLM host than any single consumer NVIDIA GPU for models in the 30-40B range.

M5 48 GB model ceiling:

Model Quantization Approx. RAM Notes
Llama 3.3 70B Q4_K_M ~40 GB Fits, leaves ~8 GB headroom
Qwen3.5 32B Q5_K_M ~22 GB Comfortable, excellent quality
Llama 3.2 11B Q8_0 ~12 GB Near full precision
Llama 3.2 3B Q8_0 ~3.5 GB Fast; good for local agents

CLI launch with Metal: 

lms server start --port 1234 --gpu-layers auto
The auto flag lets LM Studio calculate the optimal number of layers to offload to Metal based on available unified memory.

In the Settings → GPU panel, ensure the Apple Metal or llama.cpp Metal backend is selected. Use the GGUF model filter when browsing — MLX-format models require the separate MLX backend available in LM Studio 0.3.6+.

When to use it

  • When you want the fastest path to trying local models
  • When you need a simple local server for app development or evaluation

When not to use it

  • When you need multi-user, server-grade inference
  • When you already operate Ollama or vLLM successfully

Getting started

  1. Download and install LM Studio from lmstudio.ai.
  2. Open the app and search for a model (e.g., Meta-Llama-3-8B-Instruct-GGUF).
  3. Click "Download" on the desired version.
  4. Go to the "AI Chat" tab to interact with the model immediately, or the "Local Server" tab to start an API.

CLI examples

The lms CLI is bundled with the desktop application (requires version 0.4.0+).

# Check status and loaded models
lms status

# Search for and download a model
lms get meta-llama-3-8b

# Start the local OpenAI-compatible API server
lms server start --port 1234

API examples

LM Studio provides an OpenAI-compatible local server.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lmstudio")

response = client.chat.completions.create(
    model="meta-llama-3-8b",
    messages=[{"role": "user", "content": "Explain quantum entanglement."}]
)
print(response.choices[0].message.content)

Licensing and cost

  • Open Source: No
  • Cost: Free desktop app
  • Self-hostable: Local desktop runtime only

Sources / References

Contribution Metadata

  • Last reviewed: 2026-06-03
  • Confidence: high