How to run a local GGUF model with llama-cli

A local GGUF file is the direct input that llama.cpp loads for terminal inference. llama-cli provides the terminal path for checking that the file opens, the model evaluates a prompt, and text comes back before the same model is used by a server or application.

A non-interactive smoke test needs the model path, a prompt, and a token budget. --single-turn exits after the first answer, which matters for chat-template models that otherwise keep the terminal open for another prompt.

Start with a small model when checking a new build, then use the production quantization planned for the workload. A successful run shows the loaded model path, returned text, timing line, and clean Exiting… line.

1. Run a single-turn prompt against the local model file.

   $ llama-cli -m models/tinygemma3-Q8_0.gguf \
     -p "Write one short sentence about local AI." \
     -n 32 \
     --no-display-prompt \
     --single-turn
   build      : b9859-4fc4ec554
   model      : models/tinygemma3-Q8_0.gguf
   modalities : text
   ##### snipped #####
   Local AI runs the model on this computer and returns the answer in the terminal.
   [ Prompt: 113.4 t/s | Generation: 17.7 t/s ]
   
   Exiting...

   -m selects the GGUF file, -p supplies the prompt, -n limits generated tokens, and --single-turn makes the process return to the shell after the first response.

2. Confirm the startup output names the intended model file.

   If llama-cli reports that the model cannot be opened, fix the path or file permissions before changing prompt or sampling options.

3. Confirm generated text appears before the timing line.

   The exact sentence changes with the model, quantization, seed, and sampling options. The smoke test succeeds when a response appears and the command exits without a load or runtime error.

4. Start an interactive session when the single-turn smoke test succeeds.

   $ llama-cli -m models/tinygemma3-Q8_0.gguf
   > Say hello in one sentence.
   Hello from the local model.
   > /exit
   Exiting...

   Models with a chat template can enter conversation mode automatically. Use /exit or Ctrl+C to leave the session when finished.