jargon

Applied AI·Local and self-hosted inference

quick sizing check: a Q4 model needs roughly parameter-count-in-billions x 0.6 GB, plus room for context.

VRAM and unified memory

GPU memory bounds which models you can run: weights, KV cache and overhead must all fit. Apple Silicon's unified memory lets the GPU use ordinary RAM, which is why a 16GB M1 runs models a 16GB gaming PC with an 8GB card cannot.