qwen3-4b-thinking-2507-4bit
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 · reasoningmath · multi-step reasoning on small hardware
Alibaba 256K context Base model intelligence 11.9 AA
One command
rapid-mlx serve qwen3-4b-thinking-2507-4bit
Weights (2.3 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.
Point a client at it. OpenAI SDK: base_url="http://localhost:8000/v1",
model="qwen3-4b-thinking-2507-4bit". Claude Code:
ANTHROPIC_BASE_URL=http://localhost:8000 (no /v1 suffix) —
see the Claude Code guide.
Measured on real hardware
Headline machine · Mac mini M4 Pro · 48 GB
All measured machines
| Machine | Decode | First token | Peak memory | Cold boot | Weights |
|---|---|---|---|---|---|
| Mac mini M4 Pro · 48 GB rapid-mlx 0.15.5 · 2026-10-03 raw data | 93.0 tok/s | 2.9 s | 16.2 GB | — | 2.3 GB |
Each row cites its own run; numbers are never copied across machines. Mac mini M4 Pro · 48 GB: rapid-mlx 0.15.5, 2026-10-03 — raw data + method — 4 concurrent streams: 198.8 tok/s aggregate.
Will it fit your Mac?
Measured 16.2 GB Metal peak (2.7 GB process RSS) on a 48 GB M4 Pro across the whole benchmark run — up to a ~16K-token prompt and four concurrent streams, including the engine's prefix cache, which defaults to 20% of RAM. A short chat needs less; long contexts grow the KV cache, so leave headroom.
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 & profiles
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 16.2 GB Metal peak (2.7 GB process RSS) on a 48 GB M4 Pro across the whole benchmark run — up to a ~16K-token prompt and four concurrent streams, including the engine's prefix cache, which defaults to 20% of RAM. A short chat needs less; long contexts grow the KV cache, so leave headroom.
How fast is Qwen3 4B Thinking (2507) on Apple Silicon?
We measured 93.0 tok/s on a Mac mini M4 Pro · 48 GB, rapid-mlx 0.15.5, 2026-10-03; peak memory 16.2 GB on that run. Every row on this page keeps its own chip, version and date — no number is copied across machines.
How do I run Qwen3 4B Thinking (2507) locally?
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 for Aider, Codex and any OpenAI client, plus the Anthropic API at localhost:8000 for Claude Code.