Ollama: model requires more system memory than is available

Different from VRAM OOM — this is system RAM. Ollama needs to load the model file into RAM before transferring to VRAM (mmap'd or copied). On systems with low RAM and large models, the load step fails before the GPU is even involved.

Check current system RAM usage:

# macOS / Linux
free -h
# Windows PowerShell
Get-Counter '\Memory\Available MBytes'

Free system RAM by closing other apps (browsers, IDEs, Slack — these can chew 4-8 GB easily).

Use a smaller quantization so the file is smaller on disk and in RAM:

# Q4 instead of Q8 — half the file size
ollama pull llama3.1:8b-instruct-q4_K_M

Add swap (Linux):

sudo fallocate -l 16G /swapfile
sudo chmod 600 /swapfile
sudo mkswap /swapfile
sudo swapon /swapfile

Loading is slower but possible.

Get more RAM. For local AI work, 32 GB RAM is the practical floor; 64 GB or 128 GB unlocks larger models with CPU offload.