Search

Both, for different reasons. A model must fit in the memory the GPU can reach: on discrete cards that is VRAM (8GB to 24GB in this field), and our Llama 2 13B run DNF’d on 8GB cards because it wanted about 12GB. Unified and shared memory designs flip the equation: AMD’s Ryzen AI Max+ can assign up to 96GB of system RAM to the GPU, and Intel’s shared LPCAMM2 pools reach 64GB, so far larger models load at all, just at lower speeds. Fit determines whether a model runs; the silicon determines how fast.

Inference is one of the heaviest sustained loads a laptop can run, so plan on wall power for real sessions. The battery numbers on this page are PCMark 10 Modern Office results, a productivity measure, and the spread is enormous: 3 hours 39 minutes on our fastest AI laptop versus nearly 24 hours on the shared-memory Dell. Our Laptop Battery Life Leaderboard ranks the full field.

As a working rule, a 70B model at 4-bit quantization needs roughly 40-48GB of model-accessible memory before context; comfortable interactive use with meaningful context wants more. That is why 128GB unified-memory appliances handle 70B-class models well, and why 24-32GB single-GPU systems do not make this page.

If the machine will live on a desk, yes, probably. Deskside systems offer more memory per dollar, better sustained thermals, and no battery compromise; our Best Desktops for Local AI leaderboard starts at roughly the price of the midrange laptops here. The laptops on this page are for people whose AI workload has to travel.