featuresJuly 16, 2026· 4 min read

Local, offline embeddings with CEMBED (no external API)

Embeddings usually mean an API key and a per-token bill. CEMBED turns text into vectors using the bundled local model, for free, with nothing leaving your machine.

Embedding text is table-stakes infrastructure that quietly costs money and leaks data: most teams call a hosted embeddings API, paying per token and shipping their text to a third party. CEMBED removes both costs.

It embeds text with the bundled crowsight model (384-dimensional) in-process, with no external API and no egress, and repeated text is served from an embedding micro-cache so the second call is effectively free. On its own quality bench, the bundled embedder hits 100% recall at a 1% or lower false-positive rate on genuine rephrasings, ties the hosted API models, and beats other open models.

crowkis cli

CEMBED "explain vector caches"   # -> a 384-dim vector, local, free

In plain words: The cheapest embedding is the one you already computed; the second cheapest is the one that never left your machine. CEMBED is both.

You can also bring your own embedder or reranker, any sentence-transformers MiniLM or GTE export works via ONNX, so you're never locked to the default. Having a free local embedder as a primitive is more useful than it sounds: retrieval, dedup, clustering, and your own semantic features can all share one consistent, no-bill vector source.