EN
3 November 2026 · Jakub M.
MacBook Pro M5 Max with 128 GB: how far a laptop gets with local AI
Vše Smart ran six open models from 8 to 120 billion parameters on a MacBook Pro M5 Max with 128 GB. All of them fit without any system tweaks, gpt-oss-120b generates 110 tokens per second, and the chip draws only 37–50 W. Long contexts and large dense models are where the laptop runs out of breath.
Source: MacBook Pro M5 Max se 128 GB RAM v testu lokálních AI modelů, Ludvík Eckerman, Vše Smart, 2 November 2026 (in Czech). All figures below come from that article.
Can a laptop handle serious local AI? The Czech site Vše Smart tested the top MacBook Pro with the M5 Max chip (18-core CPU, 40-core GPU) and 128 GB of unified memory with 614 GB/s bandwidth. Six open models, two inference engines and a look at power draw.
How it was tested
- System: macOS 27.2, High Power mode, on AC power, default GPU memory limit (about 107 GB).
- Software: llama.cpp (build 11669, Metal) and MLX 0.32.4 with mlx-lm 0.32.2.
- Models: Llama 3.1 8B, gpt-oss-20b, Qwen3 30B-A3B, Qwen3 32B, Llama 3.3 70B and gpt-oss-120b, in 4-bit quantisation (gpt-oss in its native MXFP4).
- Method: contexts of 0 to 32,768 tokens, three runs each; chip power measured with powermetrics; a ten-minute sustained run with Qwen3 30B-A3B.
Results
| Model | Generation, llama.cpp | Generation, MLX | Generation at 32k context (MLX) | First token at 32k (MLX) | Peak memory (MLX) | Chip power | Tokens per watt |
|---|---|---|---|---|---|---|---|
| Llama 3.1 8B | 102 | 115 | 60 | 17 s | 9.5 GB | 49.8 W | 2.02 |
| gpt-oss-20b | 156 | 142 | 105 | 7.5 s | 13.7 GB | 48.5 W | 3.24 |
| Qwen3 30B-A3B | 155 | 152 | 72 | 15 s | 21.1 GB | 41.1 W | 3.69 |
| Qwen3 32B | 26 | 28 | 18 | 72 s | 27.8 GB | 48.3 W | 0.53 |
| Llama 3.3 70B | 12.5 | 13.6 | 10 | 132 s | 51.3 GB | 40.5 W | 0.31 |
| gpt-oss-120b | 110 | 94 | 70 | 13.5 s | 65.5 GB | 37.3 W | 2.92 |
Generation in tokens per second, 4-bit models (gpt-oss in MXFP4), empty context unless stated. Chip power during generation in llama.cpp, chip only (CPU, GPU, Neural Engine). Source: Vše Smart, published 2 November 2026.
Everything fits into 128 GB
All six models ran without raising the memory limit. The largest, gpt-oss-120b, peaked at 65.5 GB, leaving plenty of room. Thanks to its mixture-of-experts design it still generates 110 tokens per second in llama.cpp.
Mixture of experts wins, dense models struggle
The fastest models were those that use only part of their parameters per token: Qwen3 30B-A3B and gpt-oss-20b at about 155 tokens per second. The dense Llama 3.3 70B managed 12.5–13.6 tokens per second, and with a 32,000-token context it dropped to around 10, while the first token took 132 seconds. For long documents with large dense models, the laptop is at its limit.
Efficient and steady
During generation the chip drew 37–50 W and about 2.1 W at idle. The most efficient model was Qwen3 30B-A3B with 3.69 tokens per watt. In ten minutes of continuous load its speed stayed at about 152 tokens per second, with no sign of throttling. The figures cover the chip only, not the whole laptop, and the test ran on AC power.
When 128 GB is not enough
The MacBook Pro M5 Max is an excellent portable machine for models up to about 100 GB. Larger models such as Qwen3 235B, which needs about 133 GB, do not fit at all, and with 70B-class dense models and long contexts it becomes slow. In our clash of the titans, the Mac Studio M5 Ultra was 2.5 to 2.8 times faster with these models.
As the article notes, the test measured speed, not the quality of answers.
Need more?
Rent a Mac Studio M5 Ultra with 256 GB from one day and run the models that don’t fit on a laptop.