AP Apple Prosumer Unified memory

Apple M1 Ultra 64GB

Runs downloadable LLMs up to ~95B at Q4, ~121.4-146.4 tok/s on an 8B model.

Other configs: 128GB
Apple M1 Ultra 64GB
64GB
Unified memory
800
GB/s bandwidth
95B
Max model (Q4)
$3,700
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 90.7-109.3 t/s
≤764k ctx · then ~1.1-1.4 (slow)
181.3-218.7 t/s
≤810k ctx · then ~2.3-2.7 (slow)
323.8-390.5 t/s
≤830k ctx · then ~4-4.9 (slow)
7-8B 34-41 t/s
≤306k ctx · then ~0.4-0.5 (slow)
68-82 t/s
≤367k ctx · then ~0.9-1 (slow)
121.4-146.4 t/s
≤394k ctx · then ~1.5-1.8 (slow)
13B 20.9-25.2 t/s
≤184k ctx · then ~0.3-0.3 (slow)
41.8-50.5 t/s
≤263k ctx · then ~0.5-0.6 (slow)
74.7-90.1 t/s
≤298k ctx · then ~0.9-1.1 (slow)
32B won't run 17-20.5 t/s
≤92k ctx · then ~0.2-0.3 (slow)
30.4-36.6 t/s
≤146k ctx · then ~0.4-0.5 (slow)
70B won't run won't run 13.9-16.7 t/s
≤52k 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 ~57.6 GB usable. Beyond "ctx", KV spills to SSD swap (rough degraded speed shown).

Run popular models on the Apple M1 Ultra 64GB

Videos about the Apple M1 Ultra 64GB

Is M1 Ultra enough for MACHINE LEARNING? vs RTX 3080ti

NO other PC can match it for $4,000? - M1 Ultra Mac Studio review

More hardware like the Apple M1 Ultra 64GB

About the Apple M1 Ultra 64GB

Apple M1 Ultra 64GB is a unified-memory system from Apple on Apple M1, 2022. For local genAI the numbers that matter are its 64 GB (how big a model + context fits) and 800 GB/s bandwidth (generation speed).

Frequently asked questions

What size LLM can the Apple M1 Ultra 64GB run?

With ~57.6 GB usable it runs up to ~95B at Q4 (~26.5B at FP16) at 8k context. Bigger spills into SSD swap and slows sharply.

How many tokens/sec does the Apple M1 Ultra 64GB generate?

Generation is bandwidth-bound. At 800 GB/s an 8B model at Q4 runs ~121.4-146.4 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 57.6 GB it spills to SSD swap and throughput collapses. Each row shows the max usable context.