AI Potluck
Contribute
Back to Gap Map Model components / Embeddings & retrieval models

mxbai-embed

Mixedbread
open weights / Overall score: 3.3

mxbai-embed is Mixedbread's open-weight English sentence embedding line, headed by mxbai-embed-large-v1, a BERT-large bi-encoder with a 512-token window, released in March 2024. The family also includes a 2D-Matryoshka variant, an xsmall checkpoint and a German model built with deepset. All of them support Matryoshka truncation and binary quantization, which is the line's distinguishing feature: a 1024-dimension float vector can be cut to a fraction of its storage cost with most of its retrieval quality intact.

Scored on the embed line only; Mixedbread's mxbai-rerank and mxbai-colbert checkpoints are separate products and are excluded from the adoption sum. Verified 2026-09-11 via the mxbai-embed-large-v1 model card and the four Hub API records.

Openness

3 high confidence
3.0
weights
open(four ungated Apache-2.0 checkpoints)
data
closed(training data not published
code
partial(inference, Matryoshka truncation and binary-quantization snippets on the card)
license
Apache-2.0(OSI)

Apache-2.0 and ungated on every checkpoint in the line. The corpus is the gap: the card's only claim about training data is that it does not overlap MTEB, which is a statement about evaluation hygiene rather than a description of the mixture, and no dataset or training code ships. Open weights, closed data, the ladder's fallthrough rung.

Adoption

4 high confidence
4.0

3,113,938 downloads in the trailing 30 days across the four embed checkpoints - 2,807,219 for mxbai-embed-large-v1, 168,988 for the 2D-Matryoshka variant, 113,162 for the deepset German model and 24,569 for xsmall. Worth reading against the capability band: this is a 2024 English model still pulling millions of downloads a month, which is what an entrenched default looks like after the frontier has moved past it. Mixedbread's rerank and colbert lines add roughly another 800k and are excluded as separate products.

Capability

3 high confidence
3.0

Rung 3, a superseded generation with a narrower scope - the rung's two conditions both hold rather than one. The line is English-only with a 512-token window, and its published number is an MTEB v1 average from March 2024 that has not been refreshed against v2, so it cannot be compared with the anchor on the same instrument at all. Placed level with granite-embedding rather than against Harrier-OSS: Granite R2 is the nearest peer that is also a compact bi-encoder with published numbers, and it is a generation newer with sixteen times the context. Still useful, and the download figures say people use it, but the evidence for the band is a two-year-old leaderboard.

Verified 2026-09-11