First run¶

Inspect your model¶

spum info /path/to/model.gguf

A GGUF extension alone does not guarantee that an architecture or quantization is supported. Use the model's distribution instructions and the current compatibility information.

Generate a short answer¶

spum run model.gguf --gpu \
  --prompt "Explain why an optical image can be out of focus." \
  --max-tokens 128

Use --cpu for CPU execution. --gpu selects the supported GPU path; it does not move the run to a remote GPU.

Use deterministic sampling¶

spum run model.gguf --gpu \
  -p "What is 17 times 23?" \
  --temperature 0 --seed 42 --max-tokens 32

A fixed seed and greedy sampling improve repeatability, but different files, precision settings, backends or engine versions can still change results.

Start a conversation¶

spum chat model.gguf --gpu --context-size 4096

Type /reset to clear the current conversation and /quit or /exit to leave. Set a context size appropriate for the model and available memory.

Ask for help¶

spum help run
spum serve --help
spum music --help
spum capabilities --json