Run Qwen3 4B Thinking (2507) on a Mac
The reasoning twin of Qwen3 4B Instruct: it thinks before it answers, which buys accuracy on math and multi-step questions at the cost of latency. rapid-mlx separates the thinking channel automatically.
4 B dense · 4-bit · reasoning math · multi-step reasoning on small hardware
One command
$ rapid-mlx serve qwen3-4b-thinking-2507-4bit
Weights (2.1 GB) download on first run; you get an OpenAI-compatible
endpoint at http://localhost:8000/v1. No rapid-mlx yet? It's one line —
curl -fsSL https://rapidmlx.com/install.sh | bash — or take the
desktop app.
Measured on real hardware
| Machine | Decode | First token | Peak memory | Cold boot | Weights |
|---|---|---|---|---|---|
| Mac mini M2 Pro · 32 GB | 61.3 tok/s | 0.41 s | 3.0 GB | 3.9 s | 2.1 GB |
Measured on rapid-mlx 0.12.10, 2026-08-11. Decode: median of 3 runs, temperature 0, 256-token saturating generation, engine-reported token counts, unique salt per request (no prefix-cache hits). TTFT: median of 3, short prompt. Boot: process spawn to first completed token. Peak RSS: 0.5 s sampling across the run.
Will it fit your Mac?
Peak resident memory measured 3.0 GB during a 256-token generation. The KV cache grows with context length, so treat that as a floor, not a ceiling. For the conservative install-default placement see the hardware tiers table; to compare against every model your RAM can hold, use the live picker.
Variants & alternatives
qwen3-4b-instruct-2507-4bit— The instruct variant — direct answers, lower latency
FAQ
How much memory does Qwen3 4B Thinking (2507) need on a Mac?
Measured peak resident memory was 3.0 GB on rapid-mlx 0.12.10 during a 256-token generation (M2 Pro, 32 GB). The weights are 2.1 GB on disk. Longer contexts grow the KV cache beyond this, so leave headroom.
How fast is Qwen3 4B Thinking (2507) on Apple Silicon?
We measured 61.3 tokens/sec sustained decode and 0.41 s time-to-first-token on a Mac mini M2 Pro (32 GB), median of 3 runs at temperature 0.
How do I run Qwen3 4B Thinking (2507) locally?
Install rapid-mlx (curl -fsSL https://rapidmlx.com/install.sh | bash, or brew install rapid-mlx), then: rapid-mlx serve qwen3-4b-thinking-2507-4bit — the weights download on first run and you get an OpenAI-compatible endpoint at localhost:8000/v1 that works with Cursor, Claude Code, Aider, and any OpenAI client.