A CPU-only Ollama model selection cheat sheet by memory

On 2026-11-06, the AI section of DEV Community published a cheat-sheet style article explaining how to pick a model for running Ollama on a server without a GPU. The method it describes is simple: use the download size listed for each tag on ollama.com as the baseline for how much RAM the model will need before the context is added, then keep the model file at roughly half of the physical memory under the default context so there is still room left for the context window and the OS, and finally confirm the actual footprint on the server with the ollama ps command.

As worked examples, the article lists two tags with their download sizes and context windows: gemma3:270m at 292 MB and a 32K context window, and qwen3:0.6b at 523 MB and a 40K context window. It also notes two practical caveats for the CPU-only case: a 27B model needs more than 16 GB of memory, and on the CPU the larger the model the slower it runs, so sizing should be validated with real tests rather than assumed. That is the full extent of what the available material covers, and there is no follow-up report yet on any further development of this topic.