New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
Apple M3 Max 36GB
Runs downloadable LLMs up to ~50B at Q4, ~45.5-54.9 tok/s on an 8B model.
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 size | FP16 | Q8 | Q4 |
|---|---|---|---|
| 3B |
34-41 t/s ≤380k ctx · then ~1.1-1.4 (slow) |
68-82 t/s ≤426k ctx · then ~2.3-2.7 (slow) |
121.4-146.4 t/s ≤446k ctx · then ~4-4.9 (slow) |
| 7-8B |
12.8-15.4 t/s ≤114k ctx · then ~0.4-0.5 (slow) |
25.5-30.8 t/s ≤175k ctx · then ~0.9-1 (slow) |
45.5-54.9 t/s ≤202k ctx · then ~1.5-1.8 (slow) |
| 13B |
7.8-9.5 t/s ≤30k ctx · then ~0.3-0.3 (slow) |
15.7-18.9 t/s ≤109k ctx · then ~0.5-0.6 (slow) |
28-33.8 t/s ≤144k ctx · then ~0.9-1.1 (slow) |
| 32B | won't run | won't run |
11.4-13.7 t/s ≤50k ctx · then ~0.4-0.5 (slow) |
| 70B | won't run | won't run | won't run |
| 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 ~32.4 GB usable. Beyond "ctx", KV spills to SSD swap (rough degraded speed shown).
Run popular models on the Apple M3 Max 36GB
Videos about the Apple M3 Max 36GB
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 36GB
About the Apple M3 Max 36GB
Apple M3 Max 36GB is a unified-memory system from Apple on Apple M3, 2023. For local genAI the numbers that matter are its 36 GB (how big a model + context fits) and 300 GB/s bandwidth (generation speed).
Frequently asked questions
What size LLM can the Apple M3 Max 36GB run?▶
With ~32.4 GB usable it runs up to ~50B at Q4 (~14.5B at FP16) at 8k context. Bigger spills into SSD swap and slows sharply.
How many tokens/sec does the Apple M3 Max 36GB generate?▶
Generation is bandwidth-bound. At 300 GB/s an 8B model at Q4 runs ~45.5-54.9 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 32.4 GB it spills to SSD swap and throughput collapses. Each row shows the max usable context.