M2 Max 96GB vs M3 Pro 18GB for local LLMs
7 reports on the M2 Max 96GB and 1 on the M3 Pro 18GB
Which is faster for local LLMs?
On Qwen3 8B at 4-bit, the M2 Max 96GB runs at 43.0 tokens per second (one run) and the M3 Pro 18GB at 25.2 (one run). Those are community reports whose engines and settings differ, so the gap is a rough sign, not a controlled measure of the two cards. Memory bandwidth, which caps how fast a card can write, is 2.7x higher on the M2 Max 96GB (410 GB/s on the M2 Max 96GB, 154 on the M3 Pro 18GB).
- Memory
- 96GB unified
- Memory bandwidth
- 410 GB/s
- FP16 compute
- 13.6 TFLOPS
- Reports
- 7
- Memory
- 18GB unified
- Memory bandwidth
- 154 GB/s
- FP16 compute
- 7.4 TFLOPS
- Reports
- 1
Measured on both
Median tokens per second from plain runs (one device, one request, no speculative decoding, the whole model in the device's memory) of the same model size at the same quant level. Engines and context lengths can still differ between the runs. How to read these
| Model | M2 Max 96GB | M3 Pro 18GB |
|---|---|---|
| Qwen3 8B4-bit | 43.0one run | 25.2one run |
Estimated by the calculator
The quant the calculator recommends for each card and its estimated speed at a 32,768-token context with 32 GB of system RAM, from memory bandwidth and the model's shape. Where a model was also measured above, the measurement wins.
| Model | M2 Max 96GB | M3 Pro 18GB |
|---|---|---|
| Qwen3 8B | 17.6 t/sFP16 | 33.8 t/sQ3_K_M |
| Qwen3.8 27B | 5.2 t/sFP16 | doesn't fit at these settings |
| Qwen3.8 125B · 6B active | 36.2 t/sQ3_K_M | doesn't fit at these settings |
| DeepSeek V4 Flash 284B · 13B active | doesn't fit at these settings | doesn't fit at these settings |
| Qwen3.6 35B · 3B active | 37.6 t/sQ8_0 | doesn't fit at these settings |
| Qwen3.6 27B | 5.2 t/sFP16 | doesn't fit at these settings |
| Gemma 4 26B · 4B active | 12.5 t/sFP16 | 35.1 t/sQ2_K |
Change the context, RAM or model in the calculator: M2 Max 96GB · M3 Pro 18GB
Every report on each card, with its source: M2 Max 96GB · M3 Pro 18GB