AP Apple Consumer Unified memory

Apple M3 Max 48GB

Runs downloadable LLMs up to ~69.5B at Q4, ~60.7-73.2 tok/s on an 8B model.

Other configs: 36GB · 64GB · 128GB
Apple M3 Max 48GB
48GB
Unified memory
400
GB/s bandwidth
69.5B
Max model (Q4)
$3,200
MSRP

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to SSD swap and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 45.3-54.7 t/s
≤545k ctx · then ~1.1-1.4 (slow)
90.7-109.3 t/s
≤591k ctx · then ~2.3-2.7 (slow)
161.9-195.2 t/s
≤611k ctx · then ~4-4.9 (slow)
7-8B 17-20.5 t/s
≤196k ctx · then ~0.4-0.5 (slow)
34-41 t/s
≤257k ctx · then ~0.9-1 (slow)
60.7-73.2 t/s
≤284k ctx · then ~1.5-1.8 (slow)
13B 10.5-12.6 t/s
≤96k ctx · then ~0.3-0.3 (slow)
20.9-25.2 t/s
≤175k ctx · then ~0.5-0.6 (slow)
37.4-45.1 t/s
≤210k ctx · then ~0.9-1.1 (slow)
32B won't run 8.5-10.3 t/s
≤37k ctx · then ~0.2-0.3 (slow)
15.2-18.3 t/s
≤91k ctx · then ~0.4-0.5 (slow)
70B won't run won't run 6.9-8.4 t/s
≤8k ctx · then ~0.2-0.2 (slow)
120B won't run won't run won't run
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (Apple). Cyan = fits ~43.2 GB usable. Beyond "ctx", KV spills to SSD swap (rough degraded speed shown).

Run popular models on the Apple M3 Max 48GB

Videos about the Apple M3 Max 48GB

Space Black M3 Max MacBook Pro Review: We Can Game Now?!

Apple M3 Max MacBook Pro - A Long Term User Review

More hardware like the Apple M3 Max 48GB

About the Apple M3 Max 48GB

Apple M3 Max 48GB is a unified-memory system from Apple on Apple M3, 2023. For local genAI the numbers that matter are its 48 GB (how big a model + context fits) and 400 GB/s bandwidth (generation speed).

Frequently asked questions

What size LLM can the Apple M3 Max 48GB run?

With ~43.2 GB usable it runs up to ~69.5B at Q4 (~19.5B at FP16) at 8k context. Bigger spills into SSD swap and slows sharply.

How many tokens/sec does the Apple M3 Max 48GB generate?

Generation is bandwidth-bound. At 400 GB/s an 8B model at Q4 runs ~60.7-73.2 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 43.2 GB it spills to SSD swap and throughput collapses. Each row shows the max usable context.