DEV Community

Nerav Doshi
Nerav Doshi

Posted on

Deployed llama3.2:1b locally via Ollama

Context: Ollama is a tool for running large language models directly on your own machine, instead of calling a hosted API. llama3.2:1b refers to Meta's Llama 3.2 model at 1 billion parameters — "parameters" being roughly the number of tunable values inside the model that determine what it's learned. 1 billion is small by LLM standards (production-grade models often run into the hundreds of billions), which is exactly why it's a good fit for testing on a laptop or Mac mini: fast to download, fast to load, and light enough to actually watch its resource usage in real time.

Ran: Installed Ollama, pulled llama3.2:1b, and loaded it with a prompt to force it into memory.

Result: ollama ps reported 1.5 GB for the loaded model. Breaking that down with ps aux, the actual llama-server process held 1.24 GB RSS, with the background ollama serve daemon adding a separate ~22 MB. The ollama ps figure (1.5 GB) runs a bit higher than raw process RSS — likely accounting for GPU-resident memory that doesn't show up the same way in ps. Worth noting: this was captured after the model was already loaded, not before, so it's a snapshot rather than a true before/after delta.

Takeaway: ~1.25 GB of process RSS plus GPU memory overhead is the real number to size a resources.requests.memory floor around for this model — not the ~1.3 GB download size, which undercounts the actual runtime footprint.

$ ollama ps
NAME           ID              SIZE      PROCESSOR    CONTEXT    UNTIL
llama3.2:1b    baf6a787fdff    1.5 GB    100% GPU     4096       About a minute from now

$ ps aux | grep ollama
flyers  95443  0.1  7.7  436962064  1295248  ??  S  llama-server --model ... -c 4096
flyers   1456  0.0  0.1  436899728    22336  ??  S  ollama serve
Enter fullscreen mode Exit fullscreen mode

Top comments (0)