Ollama keep_alive tested: 5-minute default, instant VRAM release

The test targets a practical gap in local AI tooling: loading a model into VRAM is visible, but the unload schedule is not, and Ollama's own status output can mislead. Because ollama ps reports the UNTIL column as time remaining rather than the configured duration, the default 5-minute keep-alive shows up as "4 minutes from now", which the author initially mistook for a 4-minute setting. The same display quirk explains the other results: an explicit keep_alive of "30m" surfaced as "29 minutes from now", and a keep_alive of "0" produced an empty table only after a short delay.

The three scenarios map onto distinct workloads. The default suits occasional prompts; the 30-minute setting is aimed at workloads that call the model repeatedly and would otherwise pay a reload cost each time; and the "0" setting is for the end of a batch job, when the GPU should go back to other work. The author's suggested habit is to send one request with keep_alive: "0" when finished, rather than relying on the timer.

The measurement was run on Ollama 0.40.4 under Windows 11 build 10.0.26200, using qwen2.5:0.5b as a small test model that occupied 936 MB and ran 100% on GPU. Since the numbers come from one setup, users on other versions or hardware may see different timings, and the author frames the 1–2 second release lag as an observation rather than a fixed rule.