Onyx-mE5-Small (4-bit MLX)

4-bit MLX quantization of intfloat/multilingual-e5-small (MIT), for the Onyx on-device RAG embedder (OnyxEmbed). group_size 64, bits 4 on the attention/FFN dense matrices + word_embeddings; LayerNorms, biases, and positional embeddings stay fp. ~71 MB (from ~449 MB fp32).

Usage keeps e5 conventions: mean pooling, and the query: / passage: prefixes. Verified recall-neutral vs fp32 (English + cross-lingual). Inherits the base model MIT license.