CLI reference

Global flags

These work with every command.

Flag	Meaning
--json	machine-readable JSON output
--threads N	number of worker threads (default: all cores)
--timings	print how long each step took (stderr)
-q, --quiet	no progress bars or messages
--scan <dir>	also look for .gguf files in this folder (repeatable)
--ui cli\|web	interface to start when no command is given (or env NEURONDUMP_UI)

Progress bars go to stderr, so --json output can be piped safely.

Commands

list

All local models with source, architecture, parameter count, quantization, context and size.

inspect <model>

Full summary of one model.

Option
--all	inspect every model instead of one
--stats	also compute weight statistics (reads the whole file)
--full	don't shorten long arrays; include the whole vocabulary in JSON
--metadata	also print every metadata key
--out <dir>	save <model>.inspect.json there

search <text>

Searches model names, architectures and quantizations. With --model <m> it also searches that model's tokens, tensors and metadata keys. --limit N sets the number of results per kind (default 10).

neurondump search qwen
neurondump search attn_q --model llama3.2:1b

layers <model>

One row per layer: kind (attention / linear / MoE / dense), heads, sizes.

tensors <model>

Option
--layer N	only this layer
--role <r>	e.g. ffn_down, attn_q
--sort size\|type\|layer\|name\|offset
--limit N

vocab <model>

Option
-s, --search <text>	match text, raw form or id
--regex	treat --search as a regular expression
--class <c>	special reserved byte whitespace digit punctuation emoji cjk code word subword mixed
--script <s>	latin cyrillic han arabic hangul …
--lang <l>	code tokens of rust python javascript c_cpp java go sql shell
--limit N, --offset N	paging
--tokenize "<text>"	show the token ids for a text (same ids as llama-tokenize)

template <model>

The chat template as a tree. --raw also prints the source.

stats <model>

Weight statistics. --tensor <name> and --layer N (repeatable) limit the work. Results are cached.

analyze <model>

Builds a prune plan. See Analyze and slim.

Option
--corpus <files…>	text you care about; enables vocabulary trimming
--out <file>	where to write the plan (default <model>.prune.json)
--stats	compute weight statistics if not cached (head/neuron candidates)
--fast	skip the layer-similarity pass
--include-edge-layers	allow removing the first/last layer

slim <model> --plan <file> --out <name.gguf>

Option
--only <list>	candidate ids or groups: vocab layers tied quant projector metadata
--perplexity <file>	compare quality before/after with llama-perplexity

trace <model>

Needs a build with --features trace. See Trace.

Option
--prompt "<text>"	one prompt
--corpus <file>	several texts separated by blank lines; results are averaged
--top-k N	how many tokens / neurons to show per layer
--max-tokens N	cut long inputs

graph <model>

The model as a graph. See Graph and ablation.

Option
--level section\|layer\|block\|tensor\|fine	how deep to go (default layer)
--group-by section\|layer\|block\|quant\|kind\|importance	how to group nodes
--expand <id,...>	open these nodes (for example layer.7.attn for its heads)
--format tree\|json\|dot\|mermaid	output format (default tree)
--out <file>	write to a file

node <model> <id>

What a node contains, its statistics, its connections and whether it can be removed.

ablate <model> --cut <id,...>

Impact of removing or zeroing these nodes, then a simulated run if the build has trace.

Option
--static	only the instant estimate, no inference
--prompt "<text>"	prompt to compare (default "The capital of Italy is")
--corpus <file>	texts separated by blank lines: perplexity and mean KL
--tokens N	greedy tokens to generate (default 16)
--force	allow breaks_model cuts
--save-plan <file>	write the cut as a prune plan for slim

sweep <model> --corpus <file>

Cuts each layer (--level layer) or each head of one layer (--level head --layer N) in turn and ranks them by damage. --out writes the JSON, and --save-plan <file> --top N writes the N least damaging as a prune plan.

compare <model> <model> [...]

Side by side: sizes, layers, tokenizer, vocabulary overlap, differing metadata.

export <model>

Option
--format json\|ndjson\|csv
--table tensors\|vocab\|metadata\|layers	for ndjson / csv
--out <file>	default: stdout

config

config show, config set <key> <value>, config reset, config path. Keys: default_ui, port, scan_dirs, output_dir, threads.

web

--port N, --open, --host. See Web interface.

Interactive mode (REPL)

neurondump cli

Pick a model from the list (type to filter). The prompt then shows it:

neurondump[llama3.2:1b]> info
neurondump[llama3.2:1b]> layer 3
neurondump[llama3.2:1b]> vocab Rome --limit 5
neurondump[llama3.2:1b]> stats &          # & = run in the background
neurondump[llama3.2:1b]> jobs
neurondump[llama3.2:1b]> use nomic-embed-text

Command
use <model>	switch model
search <text>	search models + the current model's tokens, tensors, metadata
models	list models
info, layers, layer N, tensors, vocab, template, stats	same as the commands above, for the current model
analyze, slim, trace <prompt>, compare <model>, export	same as above
graph [options], node <id>	the graph of the current model
cut <id>..., uncut <id>\|all, selection, impact	build a cut selection and see its static impact
simulate [prompt] [--tokens N] [--corpus f]	run the selection in memory (& for background)
sweep --corpus f [--level head --layer N]	rank layers or heads by damage (& for background)
plan save <file>	write the selection as a prune plan
jobs, job <id>, cancel <id>	background jobs
web	open the web interface on the same session
rescan	look for models again
help, exit

Tab completes commands, model names, tensor names, token classes and graph node ids. History is kept between sessions. Long output is shown one page at a time.

Exit codes

Code	Meaning
0	success
1	error (message on stderr)
2	no command given and no terminal (prints help)
130	cancelled