Options and defaults

These options are implemented by the CLI. Some are command-specific; accepting a flag syntactically does not make it meaningful for every command. --json, --output and --quiet are explicitly restricted to run.

Input and output

Option	Value / default	Meaning
--model	File path	Alternative to positional model file
-p, --prompt	Text	User prompt
--prompt-file	UTF-8 file, up to 64 MiB	Read prompt from disk; cannot be combined with --prompt
--system	Text	System instruction; default is a general helpful-assistant instruction
--system-file	UTF-8 file, up to 64 MiB	Read system instruction from disk; takes precedence over --system
--image	Image file	Image input for compatible run or logits use
--mmproj	Matching projector file	Explicit vision projector
--vision-cpu	Flag	CPU image encoding
--no-image-cache	Flag	Disable persistent image caching for this invocation
--raw	Flag	Bypass normal chat formatting
--parse-special	Flag	Parse special tokens for tokenizer inspection
--json	Flag	run response and timings as JSON
-o, --output	File path	run output file; existing content is replaced
-q, --quiet	Flag	Suppress run timing messages; preserve errors

Generation

Option	Default	Meaning
-n, --n-predict, --max-tokens	256	Maximum newly generated tokens
--temp, --temperature	0.7	Sampling temperature; 0 selects greedy generation
--top-k	40	Top-k candidates
--top-p	0.9	Nucleus probability, from 0 to 1
--min-p	0	Minimum relative probability, from 0 to 1
--repeat-penalty	1.1	Repetition penalty; must be positive
--seed	0	Sampling seed
-c, --ctx, --context-size	0	Context size; 0 uses the model/runtime default
-t, --threads	0	CPU worker count; 0 uses the runtime default

--max-tokens must be greater than zero. Temperature must be finite and non-negative. Probability settings must be finite and within their documented range.

Execution

Option	Default	Meaning
--gpu	Off	Enable the supported GPU backend
--cpu	CPU	Explicit CPU execution; last CPU/GPU choice wins
--gpu-fp16 auto\|on\|off	auto	GPU precision policy
--compact off\|on\|auto	off	Compact mode; can change model quality, so compare against your baseline
--no-spec	Flag	Disable prompt-lookup acceleration
--spec	Flag	Explicitly enable supported prompt-lookup mode
--draft	Assistant file	Select a compatible assistant model
--no-draft	Flag	Disable assistant-model drafting
--draft-max	0	Maximum draft length; 0 uses runtime default
--draft-fixed	Flag	Disable adaptive assistant draft length
--history on\|off	off	Optional local output history
--no-pack	Flag	CPU compatibility/diagnostic switch
--profile	Flag	Additional benchmark profiling output

Precision, compact and drafting controls are workload-specific. A compatible assistant is required for assistant-model drafting. These controls do not establish task accuracy.

Server and benchmarks

Option	Default	Command
--host	127.0.0.1	serve bind address
--port	11436	serve port
--n-prompt	128	bench synthetic prompt length
--reps	3	bench measured repetitions, following warm-up
-f, --file	Corpus path	ppl evaluation text
--chunks	8	ppl maximum context chunks

Link and music commands have their own Link and MIDI option tables. Internal development controls are not part of the public supported interface.