Small-batch open models.

Amaretto is a line of open-weight models, trained with care, released with full weights, and sized for the hardware people actually have.

The line

amaretto‑embed‑148m

embeddinggguf

Compact embeddings for retrieval, RAG, semantic search, clustering, and classification — EmbeddingGemma‑300m with its vocabulary trimmed to the languages it serves. Half the encoder, 2.5× less RAM, and 99.8% of the original's retrieval quality. Best for: search, RAG, clustering, classification. Eight Latin-script languages and source code.

Params
148 M
Dimensions
768
Context
2,048
Base
EmbeddingGemma
License
Gemma