Models

mnm models lets you inspect the corpus's embedding model and prime the local cache directory. Both the embedder and the reranker run remotely on VoyageAI. Nothing is downloaded, run, or cached on your machine (no Python, no ONNX, no model files, no GPU).

What runs where

Role	Model	Where it runs	Notes
General embedder	voyage-context-3	VoyageAI (remote)	Contextualized embeddings: each document's chunks are embedded together, so every chunk vector carries document-level context. 1024-dimensional by default (Matryoshka-configurable). A query is embedded as a single-chunk document.
Code embedder	voyage-code-3	VoyageAI (remote)	A second vector on code chunks. At query time code_mode (on/off/exclusive) decides whether this code-vector list joins the RRF fusion. Forced off for mode=fts.
Reranker	rerank-2.5 (or rerank-2.5-lite)	VoyageAI (remote)	Used when reranking is requested (on by default). rerank-2.5-lite is lower latency and billed at half tokens server-side.

The corpus advertises its active embedding model as name@revision (e.g. voyage-context-3@1). If the corpus rolls the model forward, clients are told to re-embed against the new model rather than return mis-scored results.

mnm models pull

Ensures the local model-cache directory exists. Because both models are remote VoyageAI, nothing is fetched; this subcommand only primes the directory.

mnm models pull
mnm models pull --cache-dir /path/to/cache
mnm models pull --json

Flag	Notes
--cache-dir <path>	Override the local cache directory. Precedence: this flag > config [models].cache_dir > MIDNIGHT_MANUAL_MODEL_CACHE_DIR env > XDG_DATA_HOME/midnight-manual/models > HOME/.local/share/midnight-manual/models.

mnm models active

Fetches the corpus's currently active embedding model from the server. Use this to verify that your local configuration matches what the corpus is embedded with.

mnm models active
mnm models active --json

Example output:

corpus active embedding model:
  wire id:   voyage-context-3@1
  name:      voyage-context-3
  revision:  1
  dim:       1024
  provider:  voyageai

The wire_id field in --json output is the name@revision string that labels search requests. The CLI resolves this automatically at search time (a GET /v1/models/active call), so you only need this subcommand when diagnosing embedding-model mismatch errors.

Embedding paths

Which account processes your query text depends on whether you supply a Voyage key:

Path	When it applies	What happens
BYOK	VOYAGE_API_KEY set (or --voyage-api-key)	Your client embeds directly against your own Voyage account. Reranking also runs locally under your account.
Server-proxy	no Voyage key	Your client POSTs query text to the hosted server's /v1/embeddings, which calls Voyage under the operator's platform account.

Either way the query text reaches Voyage; the only question is whose Voyage account processes it. There is no on-device embedding path.