Rapid-MLX vs Ollama vs LM Studio on Apple Silicon

Three good ways to run local models on a Mac. Here's where each one fits — plus a reproducible benchmark so you can fill in the speed numbers on your own machine, not take ours on faith.

If you run local models on a Mac, you've probably landed on one of three tools: Ollama, LM Studio, or Rapid-MLX. All three are good. They optimize for different things, so the honest answer to "which is fastest / best?" is it depends on your Mac, your model, and what you're wiring it into. This page lays out where each one fits, shows the rapid-mlx numbers we've actually measured — and then hands you a benchmark you can run yourself, because a speed number we recorded on our Mac Studio won't predict what you'll see on yours.

Where each one fits

Read that as three honest lanes, not a scoreboard:

• Ollama is the one to keep if you need Windows or Linux, or a specific model that isn't in the rapid-mlx catalog. Its library is larger and it's the de-facto standard — that's a real advantage, not a footnote.
• LM Studio is the pick if you want a polished GUI to browse, download, and chat with models without touching a terminal. It also runs on Windows/Linux.
• Rapid-MLX is built narrowly for Apple Silicon: pure MLX kernels (no llama.cpp fallback), and it's the only one of the three that speaks the Anthropic API (/v1/messages) as well as OpenAI — which is what lets Claude Code point at a local model. Apache-2.0, CLI-and-server, Mac-only.

If you're on a Mac and wiring a local model into a coding agent, rapid-mlx is the tightest fit. If you need to leave the Mac, or you want a GUI, the other two are genuinely better for that.

How fast is Rapid-MLX? (measured)

Here are decode speeds we've actually recorded, single-stream, on a Mac Studio:

Run the comparison yourself

Here's the part most comparison posts skip: we haven't run a same-machine head-to-head against Ollama and LM Studio, so we're not going to print a "2–4× faster" bar we can't reproduce. Cross-engine speed swings with the exact model, quant, prompt length, and your Mac's thermal state — a number from our Studio tells you almost nothing about your M2 Air. The only comparison that's true for you is the one you run on your Mac. It takes about five minutes.

Pick the closest matching model in each tool (they won't be byte-identical quants — that's fine, note it), use the same prompt, and run each twice, reading the second (warm) run.

Rapid-MLX — the canonical single-stream benchmark:

rapid-mlx bench qwen3.5-9b-4bit
# prints decode tok/s, prefill tok/s, and TTFT.
# add --submit to publish your row to rapidmlx.com/performance (CC0)

Ollama — --verbose prints the eval rate:

ollama run qwen2.5:7b --verbose "Write a 200-word explanation of how a transformer works."
# read the "eval rate: NN tokens/s" line it prints after the response

LM Studio — load a comparable model, then either read the tok/s readout in the chat UI after a generation, or start its local server and time a request. The lms CLI also exposes model info if you prefer the terminal.

Now you have three real numbers, on one machine, for one model — the honest version of "which is fastest." If you get a reproducible result, send it our way — we'll fold community comparisons into the performance page.

Which should you pick?

• Mac + a coding agent (Claude Code, Cursor, Codex) + you want speed → Rapid-MLX. The Anthropic API surface and pure-MLX kernels are exactly what this workflow needs, and rapid-mlx launch claude-code wires it up in one command.
• You need Windows/Linux, or a model outside the catalog → Ollama. Don't fight it — it's the right tool for cross-platform and breadth.
• You want a GUI to browse and chat, no terminal → LM Studio.

And the boats-unburned option: the servers listen on different ports (rapid-mlx :8000, Ollama :11434), so you can run them side by side and move one tool over at a time. Nothing forces the choice.

Where to go next

• Moving from Ollama to rapid-mlx — the command + model-name cheat sheet, if you're switching.
• Run Claude Code on a local model — the Anthropic-API workflow rapid-mlx is built for.
• The 17-model M3 Ultra benchmark — full methodology behind the numbers above.
• Install rapid-mlx — one curl, or Homebrew. Apache 2.0.