Device	RAM	Experience
Any M-series Mac (8GB+)	~3GB	✅ Runs everywhere
M1 MacBook Air (8GB)	~3GB	✅ Comfortable
M3 Max / Pro	~3GB	✅ Near-instant responses
iPhone 15 Pro (via MLX)	~3GB	✅ Runs on-device

This is the "works on everything" model. Load time under 5 seconds on any M-series chip.

Quick Start

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("Rapid42/Qwen3.5-4B-MXFP4")

messages = [{"role": "user", "content": "Write a bash script to batch rename EXR files."}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_dict=False
)

response = generate(model, tokenizer, prompt=prompt, max_tokens=256, verbose=True)
print(response)

CLI (instant interactive chat):

mlx_lm.chat --model Rapid42/Qwen3.5-4B-MXFP4

When to Use 4B vs Larger Models

Use Case	Recommended
Quick code snippets	✅ 4B (fast)
Complex reasoning / long-form	❌ Use 27B or 35B-A3B
Always-on background assistant	✅ 4B (low overhead)
Multilingual tasks	⚠️ 4B is decent, larger is better
RAG / retrieval + answer	✅ 4B works well with good context

Why MXFP4?

MXFP4 (Microscaling FP4) uses per-block scaling factors that preserve more precision than standard int4, while remaining natively fast on Apple Silicon via MLX. For a 4B model this means near-fp16 quality at int4 memory cost.

About Rapid42

Rapid42 builds fast, precise engineering tools — from VFX pipeline utilities to optimized ML model distributions.

→ rapid42.com · ExrToPsd · Level Careers