Web interface

neurondump web --open            # default port 7878
neurondump web --port 9000

It prints a link like:

NeuronDump web UI: http://127.0.0.1:7878/?token=8cb862630b8fe923526d6c10f055950c

Open that link. The token is new at every start, and the page doesn't work without it.

Pages

Page	What you see
Models	every local model, total disk size, filter box, sortable columns; Rescan button
Overview	summary cards, architecture, sizes, tokenizer, special tokens, Ollama settings, warnings, searchable metadata, export links
vocab	search the vocabulary, filter by class / language, page through 100k+ tokens
tensors	filter and sort tensors; click one to see its statistics
layers	a colored map of the layers (attention, linear, MoE); click a layer for its tensors
graph	the model as an interactive 3D (tower, city, force) or 2D graph: open layers down to heads, inspect, search, cut, see the direct impact, simulate, sweep, see what they contain, cut them, simulate, sweep, save a plan (see Graph and ablation)
template	the chat template as a collapsible tree
stats	statistics and histogram for one tensor, or the whole model as a job
analyze	run an analysis, tick the candidates you want, press Slim
trace	run a prompt and see each layer's prediction
Compare	tick two or more models
Jobs	running and finished jobs, live progress, cancel

Screenshots

Overview of qwen3.8:27b: summary cards, hyperparameters, components, tensor types, layer kinds, tokenizer, Ollama settings and the vision projector.

vocab: searching qwen3:8b's 151,936 tokens for "Hello", with matches highlighted.

graph: qwen3.8:27b in the 3D city view. Layer 64 (the next-token prediction layer) is selected, and the side panel shows what it contains and that removing it is risky.

The same model at section level, grouped by layer. The Body section holds 65 layers; the vision projector (green, safe) is a separate file.

trace: qwen3:8b on "The capital of Italy is". The last column is the logit lens: what the model would answer if it stopped at that layer. The final prediction is " Rome" 0.825.

Compare: qwen3.8:27b next to qwen3:8b, with their vocabulary overlap and the metadata that differs.

Long work (stats, analyze, slim, trace, ablate, sweep) runs as a job in the background, so the page stays usable. A number next to Jobs in the header shows how many are running.

Search and shortcuts

• The search box in the header finds models by name, architecture or quantization. When a model is open, it also searches that model's tokens, tensors and metadata keys. Results link straight to the right page (token → vocab, tensor → its stats, key → metadata). On the results page, the "in" menu switches which model is searched.
• Press / anywhere to jump to the search box, and Esc to leave it.
• Click a column header to sort a table. Click an id, name or path in code style to copy it.
• The System / Light / Dark switch in the header sets the theme. System (the default) follows your operating system, even when it changes.

Safety

• It listens on 127.0.0.1 only (your own computer). --host 0.0.0.0 makes it reachable from your network and prints a warning. Avoid that on shared networks.
• Every API call needs the session token.
• Models are referred to by id, never by file path. Slimmed files can only be written to the output folder (~/neurondump/slim/).

HTTP API

Send the token in the x-neurondump-token header. For links and live updates, which can't set headers, use ?token= instead.

T=<token>; H="x-neurondump-token: $T"; B=http://127.0.0.1:7878
curl -H "$H" $B/api/models

Method	Path
GET	/api/info	version, trace support
GET	/api/models	model list
POST	/api/models/rescan	look for models again
GET	/api/models/{id}	overview
GET	/api/models/{id}/metadata?full=	metadata keys
GET	/api/models/{id}/vocab?q=&regex=&class=&script=&lang=&offset=&limit=	vocabulary search
GET	/api/models/{id}/tensors?layer=&role=&q=&sort=&offset=&limit=	tensors
GET	/api/models/{id}/layers	layers
GET	/api/models/{id}/template	template trees
GET	/api/models/{id}/tensors/{name}/stats	statistics of one tensor
GET	/api/models/{id}/graph?level=&group_by=&expand=	the graph; expand (comma-separated ids) loads heads / neurons / experts on demand
GET	/api/models/{id}/nodes/{node_id}	one node: contents, stats, connections
POST	/api/models/{id}/impact	static impact: {"cut":["layer.7"],"force":false}
POST	/api/models/{id}/plan-from-ablation	prune plan from {"cut":[…]} or {"sweep_job":"<id>","top":2}
POST	/api/jobs	start a job: {"kind":"stats","model":"<id>","params":{}} → 202 {"id":…}
GET	/api/jobs	all jobs
GET	/api/jobs/{id}	one job
DELETE	/api/jobs/{id}	cancel
GET	/api/jobs/{id}/events	live progress (Server-Sent Events)
GET	/api/jobs/{id}/result	result (409 while running)
GET	/api/search?q=&model=&limit=	search models (and tokens, tensors, metadata of model)
GET	/api/compare?models=a,b	comparison
GET	/api/export/{id}?format=json\|csv\|ndjson&table=	download

Job kinds: inspect, stats, analyze, slim, trace, ablate, sweep. The web API doesn't accept file paths, so pass corpus text as corpus_text for analyze and as corpus (a list) for trace, ablate and sweep.