mxbai-embed-large

mxbai-embed-large-v1 is Mixedbread AI's flagship open text embedding model: a 335M-parameter BERT-large architecture producing 1024-dimensional embeddings, released under Apache 2.0. Mixedbread reports 64.68 average across the 56-dataset MTEB suite — a hair above OpenAI's text-embedding-3-large at 64.58 — from a model small enough to run on a laptop. It has become a popular default for local RAG stacks, chosen by several open-source projects over bge-large and e5-large for its retrieval scores at a modest footprint.

The honest picture: the 512-token maximum sequence length is the real constraint — pipelines that chunk documents into 1000-token pieces will silently truncate and lose retrieval quality, so it only fits small-chunk designs. It is English-only, and the headline MTEB numbers are vendor-reported. But for English short-chunk retrieval on commodity hardware, it remains one of the best quality-per-megabyte open options.