LM Studio on a Mac: Requirements, Real Speeds, and Running It Remotely
LM Studio is the easiest way to run LLMs on a Mac with a real interface: browse models, download a quantized build, chat — no terminal. It needs an Apple Silicon Mac and realistically 16 GB of memory for 7B–8B models. You can run it two ways: on your own Mac (free, but ties up your machine), or on a dedicated remote Mac that stays on 24/7 and doubles as a shared API endpoint for your team. This guide covers both, with measured speed numbers.
What LM Studio needs on macOS
On Apple Silicon, LM Studio runs models through two engines — a llama.cpp build with Metal acceleration for GGUF files, and Apple's MLX engine — and picks up the GPU automatically, no configuration. Memory is the real requirement:
- 16 GB — one 7B–8B model (Llama 3.1 8B, Mistral 7B, Qwen 7B) at Q4 with room for context.
- 32 GB — 14B comfortably, or a 32B at Q4 with the GUI closed to essentials.
- 64 GB — 70B-class at Q4 (Llama 3.3 70B, Qwen 72B).
- 128 GB+ — Q8 70B builds and multi-model serving.
Per-model detail is in our hardware guides.
How fast is it?
LM Studio's GGUF engine is llama.cpp underneath — the same engine we benchmark on our fleet, so the numbers transfer directly. Measured on a base M4 with 16 GB (median of 3 runs, methodology here): Llama 3.1 8B — 21.2 tokens/sec, Mistral 7B — 22.8, Qwen 2.5 7B — 22.3, Qwen 2.5 14B — 11.7. An M4 Pro roughly doubles these (273 vs 120 GB/s memory bandwidth); an M4 Max doubles them again. That is real-time chat speed for anything 14B and under.
Running LM Studio on a remote Mac
The laptop workflow breaks down when you want the model available 24/7, shared with teammates, or bigger than your laptop's RAM. The fix is running LM Studio on a dedicated Mac in the cloud:
- Build a Mac on Macyou — an M4 16 GB from $99/mo for 7B-class, an M4 Pro 64 GB from $199/mo for 70B.
- Open the built-in browser desktop (WebRTC — no VNC client to install), download LM Studio from lmstudio.ai, and use the GUI exactly as you would locally.
- For API access, enable LM Studio's server (or run
lms server startheadless) — it exposes an OpenAI-compatible endpoint on port 1234 that your whole team can share.
# LM Studio's built-in server, OpenAI-compatible:
curl http://localhost:1234/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"llama-3.1-8b","messages":[{"role":"user","content":"Hi"}]}'Prefer zero setup? Macyou's catalog stacks ship Ollama pre-configured with a public OpenAI-compatible endpoint out of the box — same models, no GUI step. Many teams use LM Studio for exploration and the catalog endpoint for production.
LM Studio or Ollama?
Both run the same GGUF models on the same engine. LM Studio gives you a GUI for discovering and comparing models, inspecting prompts, and tweaking sampling; Ollama is CLI-first and easier to automate. Speed is effectively identical. Start with LM Studio to find your model, run production on Ollama — or keep both pointed at the same machine.
FAQ
Can I use LM Studio remotely from another device?
Two ways: interact with the GUI over a remote desktop (Macyou ships a browser-based one), or talk to its OpenAI-compatible server from any HTTP client. The server is the practical option for phones, tablets, and apps.
Does LM Studio use the GPU on Apple Silicon?
Yes — Metal acceleration for GGUF builds and the MLX engine are enabled automatically on Apple Silicon; there is nothing to configure.
Is 8 GB of RAM enough for LM Studio?
Only for 3B-class models at small context, and macOS itself will fight you for memory. 16 GB is the honest minimum for 7B–8B models; that is exactly why our smallest cloud config starts there.
How much does a cloud Mac for LM Studio cost?
On Macyou: $99/mo for an M4 16 GB (7B–14B models), $199/mo for an M4 Pro 64 GB (70B-class) — fixed, no per-token fees, annual −20%. See pricing.