Reference · Information

`senkani models`

Live

List, pull, or verify Senkani's local ML model registry. The registry covers the Gemma 4 vision tiers, the MiniLM-L6 embedding model, and the openai/privacy-filter Layer-3 PII classifier. Pulls are operator-explicit — Senkani never auto-downloads.

Syntax

senkani models <subcommand> [arg]

Subcommands

Subcommand

Description

list

default

Print every registered model with status (available, downloaded, verified, broken, …) and on-disk size. Pass --json for machine-readable output.

pull <model-id>

Download the named model and run its post-install verification fixture. Surfaces a clear staged-delivery error for models whose backend is not yet wired.

verify <model-id>

Re-run the verification fixture for an already-downloaded model. Updates status to verified on success, broken on failure (lastError preserved).

Examples

$ senkani models list
$ senkani models list --json
$ senkani models pull pii-classifier-int8
$ senkani models verify minilm-l6

Notes

Operator-explicit pulls. Senkani never auto-downloads weights. Layer-3 PII detection silently no-ops when the classifier is absent (Layers 1+2 still run); operators opt in via senkani models pull.
Vision and embedding models. The Gemma 4 tiers and MiniLM-L6 install via the SenkaniApp Models pane or via the MCP server's startup path. The CLI's pull command surfaces a "use the app or MCP" message for these — the MLX-Swift container loaders live in the MCP layer and the CLI doesn't link MLX.
PII classifier (T.2a). The infrastructure is shipped — registry entry, BIOES/Viterbi decoder, CLI surface. The MLX-Swift inference backend (or GGUF/llama.cpp fallback) and the real HuggingFace download driver land in phase-t2-pii-classifier-backend-wiring; until then, senkani models pull pii-classifier-int8 on a real machine surfaces the staged-delivery message.

senkani models

Syntax

Subcommands

Examples

Notes

See also

`senkani models`