MxbAI Embed is a family of state-of-the-art sentence embedding models by mixedbread-ai. These models map text into high-quality dense vector representations suitable for semantic search, retrieval-augmented generation (RAG), clustering, and classification tasks.
All MxbAI Embed models support Matryoshka Representation Learning (MRL), which allows flexible dimensionality reduction — you can truncate the output embeddings to smaller sizes (e.g., 512, 256) while retaining strong performance. This enables a cost-performance tradeoff at inference time without retraining.
The family includes English, German, and efficiency-focused variants, ranging from ultra-compact (~24M parameters) to large-scale (~340M parameters).
Key features:
High-quality dense sentence embeddings via BERT-based architectures
Matryoshka Representation Learning for flexible embedding dimensions
2D-Matryoshka variant with both layer and dimension reduction
Optimized for semantic search, retrieval, clustering, and classification
pip install -U -q keras-hub
pip install -U -q keras
Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.
Preset Table
Preset name
Parameters
Description
Base Architecture
mxbai_embed_large_en
335,141,888
24-layer MxbAI Embed large English model. Maps sentences to 1024-dimensional dense vectors with Matryoshka representation learning. Trained on high-quality English data. Top-ranked open-source embedding model on the MTEB leaderboard.
BERT-large
mxbai_embed_2d_large_en
335,141,888
24-layer MxbAI Embed 2D-Matryoshka large English model. Maps sentences to 1024-dimensional dense vectors. Supports both embedding dimension reduction and transformer layer reduction for flexible performance-efficiency tradeoffs.
BERT-large
mxbai_embed_xsmall_en
22,713,216
6-layer MxbAI Embed extra-small English model. Maps sentences to 384-dimensional dense vectors with Matryoshka representation learning. Ultra-fast and compact, based on MiniLM-L6-v2 architecture.
