Technology · Agentic → RAG · Open Source · wiki:deep

mxbai-embed-large

mxbai-embed-large-v1 (Mixedbread) is an open English sentence-embedding model with native length 1024. It is trained so shorter prefixes remain usable under Matryoshka Representation Learning — callers truncate with truncate_dim / API dimensions (e.g. 512) without a second model. The card also documents optional binary/int8 quantization as a separate cost lever from length. It is an embedder, not a vector database: you still persist vectors in Qdrant, FAISS, pgvector, or a peer store whose collection dimension matches the chosen m. Contrast OpenAI text-embedding-3-small (default 1536) / -3-large (default 3072), which reach 1024 via the same dimensions idea on a closed API.

Why it matters here

Matryoshka truncation is easier to teach against a 1024-native OSS model than against a closed API alone. In an agentic OS / RAG corridor, mxbai is a measurable alternative to OpenAI text-embedding-3-* when you want self-hosted weights, explicit truncate knobs, and optional quantization without leaving the same method taxonomy (tech-matryoshka-truncation).

How it works

Text (with a retrieval query prompt when searching) is encoded to a 1024-d float vector. Callers may request a shorter prefix length; the leading dimensions remain the representation. Optionally quantize values for cheaper indexes. Persist vectors in a store whose collection dim matches the chosen m.

Flow

When to reach for it

  • Use when: you need an OSS embedder with documented Matryoshka truncation from native 1024, optionally with quantization.
  • Skip when: you must stay on a vendor API only, or another pack already owns embeddings in the install.
  • Prefer instead: OpenAI text-embedding-3-* via API when managed embeddings win; Sentence Transformers when you are training Matryoshka models yourself.

Limits

  • Query prompts matter for retrieval (Mixedbread documents a specific search prompt).
  • Collection dim must match m; mixing 1024 and 512 in one index fails.
  • Quantization and truncation are different levers — measure both.
  • Not a ledger or authority plane.

What we checked

Claims below are backed by science sources on disk.

Matryoshka + quant support

mixedbread mxbai-embed-large-v1 (pack mirror) · STRONG

Our model also supports Matryoshka Representation Learning and binary quantization.

Native dim / truncate API

mixedbread mxbai-embed-large-v1 (pack mirror) · STRONG

Our model also supports Matryoshka Representation Learning and binary quantization.
Research inventory

9 tags · 10 out · 11 in · 3 artifacts · 1 gaps · 0 corpus docs

Catalog tags

landscape.layer
Agentic
landscape.subcategory
RAG
license_tag
Open Source
maps.dm
absent
maps.features
0
one_liner
RAG
review.depth
science
slug
mxbai-embed-large
title
mxbai-embed-large

Artifacts

  • dm_map · absent
  • features_map · present · technologies/mxbai-embed-large/features.md
  • readme · present · technologies/mxbai-embed-large/README.md

Out · alternative_to

In · alternative_to

In · in_stack