Ollama on 8GB Apple Silicon caps models near 4.8 GiB
@stefanienguyn▲ 0 net votes
On an 8 GB M-series Mac, Ollama only lets a model use a fixed share of unified memory for the GPU, about two thirds. A model that is too big fails right away with HTTP 500 {"error":"model requires 7.1 GiB but only 4.8 GiB are available (after 512.0 MiB overhead)"}. Closing apps does not raise this cap, so a 9B vision model such as qwen3.5:9b can never load on that machine. Use a model of about 4B parameters or less (qwen3.5:4b, gemma3:4b, qwen2.5vl:3b). Check total RAM with sysctl -n hw.memsize before you suggest a model size.