Run Qwen 3.5 & Qwen 3.6 Locally on Your MacBook Pro

Qwen is the open-weights family that local coding runs on — and every size of it, from a 2 GB 4B to the 122B flagship, runs on Apple Silicon. Here's the size that fits your Mac, and the one command that serves it.

Ask anyone running LLMs locally what's actually loaded on their machine and the answer, more often than not, is some size of Qwen. The family covers everything from a 4B you can run on a base MacBook Air to a 122B mixture-of-experts flagship, the licenses are permissive, and the coding and tool-calling behavior is the best in open weights right now. It's the family we lean on hardest ourselves — the model that gates every rapid-mlx release before it ships is a Qwen.

This guide is the short path: pick the right Qwen for your Mac's RAM, serve it with one command, and point your tools at it.

The 60-second version

curl -fsSL https://rapidmlx.com/install.sh | bash   # or: brew install rapid-mlx
rapid-mlx serve qwen3.6-35b-4bit                    # pick your size below

Then talk to it — it's a standard OpenAI-compatible server on http://localhost:8000:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "default", "messages": [{"role": "user", "content": "Write a binary search in Swift."}]}'

First run downloads the weights; after that, startup is seconds. No account, no API key, nothing leaves your machine.

Qwen 3.5 vs Qwen 3.6 — which one?

Two generations, both fully supported:

• Qwen 3.6 is the current generation: a dense 27B and a

mixture-of-experts 35B-A3B that activates only ~3B parameters per token — which is why it decodes fast enough to feel interactive even on laptop-class chips. If your Mac fits one, start here.

• Qwen 3.5 is the previous generation, and it earns its place with a

much wider size range: 4B, 9B, 27B dense, a 35B-A3B MoE, and the 122B-A10B flagship. Below 24 GB of RAM, the 3.5 sizes are how you get Qwen at all.

Which size fits your Mac

Sizes below are what the weights occupy on disk (and roughly in unified memory once loaded). Leave headroom for your context window and the rest of macOS.

Not sure where you land? The hardware tier guide maps every Mac from an 8 GB Air to a 512 GB Studio to a recommendation, and models.rapidmlx.com lets you filter the whole catalog by your RAM.

Tool calling and thinking work out of the box

serve auto-configures the right chat template and tool-call parser for every Qwen alias, and grammar-constrained tool calling is on by default — so function calls come back as valid JSON, not as a paragraph that almost parses. Reasoning variants stream their chain-of-thought separately as reasoning_content on the standard OpenAI wire, so your client sees clean answers with the thinking demultiplexed.

That's the boring-but-load-bearing part: it means Qwen-on-your-Mac slots into tools that expect OpenAI semantics without any glue code.

Wire it into your editor or agent

The server speaks the OpenAI protocol, so anything with a "custom base URL" box can use it:

• Claude Code — run the CLI

agent fully local against your Qwen.

• Cursor — chat, Ctrl-K and

composer against http://localhost:8000/v1.

• Codex CLI and

Aider — both verified end-to-end against the Qwen 3.6 family in our integration matrix.

For agent work specifically, prefer qwen3.6-35b-4bit or above — multi-step tool use is where the extra capability pays for itself. We hold rapid-mlx releases until the Tier-1 agents pass a live smoke test on Qwen 3.6, so this exact pairing is re-verified on every release.

How fast is it?

We publish real numbers rather than vibes: our five-model MacBook Pro benchmark puts Qwen 3.6 through coding, reasoning, tool-calling and memory tests against four rivals, and the M3 Ultra benchmark covers the desktop end. Short version: the 35B MoE is fast enough for interactive coding on a MacBook Pro, and quality-wise it's the family to beat in open weights.

Frequently asked questions

Can I run Qwen locally on a Mac with 8 GB of RAM?

Yes — qwen3.5-4b-4bit is 2.0 GB on disk and runs comfortably on an 8 GB MacBook Air with room left for the OS. It handles chat, Q&A and light code editing well. For heavier work you'll want 16 GB and the 9B, or 24 GB+ for the current-generation Qwen 3.6 models.

What's the difference between Qwen 3.5 and Qwen 3.6?

Qwen 3.6 is the newer generation, available as a dense 27B and a mixture-of-experts 35B (3B active per token) — better quality and faster decode when your RAM fits one. Qwen 3.5 is the previous generation but spans far more sizes (4B, 9B, 27B, 35B, 122B), so it's the practical choice on smaller Macs and the only route to the 122B flagship.

Does Qwen work with Claude Code and Cursor locally?

Yes. rapid-mlx serves Qwen behind an OpenAI-compatible endpoint, and the Qwen 3.6 pairing with Claude Code, Codex, Aider and Cursor's editor surfaces is verified end-to-end as part of our release gate. Point the tool's base URL at your local server and it works — no bridge, no plugin.

Is running Qwen locally actually free and private?

The weights are open, rapid-mlx is Apache 2.0, and inference happens entirely on your Mac — no account, no per-token billing, and your prompts and code never leave the machine. The only costs are disk space for the weights and the electricity your Mac was already using.