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

mxbai-rerank

Mixedbread
open weights / Overall score: 3.7

Mixedbread's cross-encoder reranking line, distinct from their mxbai-embed bi-encoders. The v2 generation (base 0.5B, large 1.5B) is trained with a three-step GRPO, contrastive and preference-learning recipe, covers 100-plus languages and 8K-token documents, and reports 57.49 BEIR average for large-v2; the v1 generation (xsmall, base, large) predates it and still carries most of the line's downloads.

Added 2026-09-11 for the embeddings_retrieval promotion. Confirmed as a real and separate line - five mxbai-rerank checkpoints ship under their own PyPI package and GitHub repository, none of which is an mxbai-embed artifact. The mxbai-colbert and mxbai-edge-colbert late-interaction models are a third line again and are not counted here. Verified 2026-09-11 via the mxbai-rerank-large-v2 model card and the mixedbread-ai/mxbai-rerank repository.

Openness

3 high confidence
3.0
weights
open(Apache-2.0 ungated safetensors across the v1 and v2 checkpoints)
data
closed(no training data released
code
partial(the mxbai-rerank package is inference only)
license
Apache-2.0(OSI)

Apache-2.0 ungated weights with an inference package but no training pipeline and no corpus, which is this category open-weights centre of gravity. The three-step GRPO, contrastive and preference-learning recipe is described in prose and in arXiv 2506.03487 rather than shipped.

Adoption

3 high confidence
3.0

About 804K downloads in the trailing 30 days across the five mxbai-rerank checkpoints - xsmall-v1 631K, large-v2 102K, base-v1 43K, large-v1 31K, base-v2 28K. The mxbai-colbert and mxbai-edge-colbert late-interaction models are a separate line and are not counted.

Capability

4 medium confidence
4.0

Rung 4, competitive frontier: real multilingual and code coverage and a widely used inference package, but the v2 generation shipped in March 2025 and its own reported BEIR average sits below the 2025-26 frontier. Level with jina-reranker, whose v3 card reports 61.94 BEIR. The two BEIR figures are not from one harness - jina own table credits mxbai-rerank-large-v2 with 61.44 - which is why this is placed by relation rather than by arithmetic on the two numbers, and why confidence is medium.

Verified 2026-09-11