M5 Max MacBook Pro nears RTX laptop LLMs

The M5 Max looks like a real contender for on-device LLM work, but these numbers are still more teaser than verdict. Apple’s real advantage is not winning tiny-model token races, it is keeping bigger models usable once mobile GPU VRAM runs out.

• –The clearest early datapoint is about 59 tok/s on DeepSeek R1 14B in LM Studio, close to laptop RTX 5080 and 5090 results in the same roundup.
• –Reddit commenters focused on the M5 Max’s 614GB/s unified memory bandwidth, arguing that Apple closes the gap when models no longer fit cleanly inside 24GB to 32GB mobile GPU memory.
• –The missing data is prompt processing, long context, and larger 30B+ or MoE workloads, which are usually more revealing than a single small-model decode test.
• –For AI developers, the interesting angle is efficiency and memory headroom: a quieter, battery-friendly laptop that can run larger local models may matter more than topping raw tokens-per-second charts.