DEVELOPER REFERENCE · VERIFIED MODEL EVIDENCE

Compare 11 source-verified model records, repository licenses, published weight formats and dated evidence. No overall model ranking.

Evidence statusSource-first verified
Latest record check2026-11-13
Latest record full verification2026-10-11

DIRECT ANSWER · SOURCE-OBSERVED REGISTRY

What BAAI model evidence is available?

Open Model Weights currently tracks 11 field-verified BAAI model records. In this indexed sample, the recorded modalities are 8 Embeddings / features; 2 Text Classification; 1 Text embeddings. The published repository license labels include 10 mit; 1 apache-2.0. The listed weight formats include PyTorch (10), Safetensors (8), ONNX (7).

Coverage describes the model repositories indexed here, not every model released by the publisher. A checked source, an unknown field and an unchanged source revision are different states. No missing value is treated as a negative answer.

11verified model records
3recorded modalities
2distinct license labels
11model repository links

SOURCE-BACKED PUBLISHER PROFILE

BAAI: models, research and access

BAAI (Beijing Academy of Artificial Intelligence) publishes the BGE (BAAI General Embedding) retrieval model family. Its FlagEmbedding project provides official usage patterns for text embeddings, multilingual dense/sparse retrieval and pairwise reranking.

An embedding encoder retrieves candidate passages from a corpus; a reranker scores query-passage pairs after retrieval. These model roles, language coverage and scoring conventions cannot be reduced to a single 'best BGE model' without an application-specific evaluation.

01

BGE-M3: dense, sparse and multi-vector retrieval

BAAI's own FlagEmbedding documentation identifies three representation modes for BGE-M3: dense vectors, lexical sparse weights and multi-vector late interaction. It also describes multilingual training and longer documents. Actual search quality depends on the implementation, retrieval index, document language and score fusion weights.

02

BGE v1.5 English and Chinese embedding sizes

BGE's v1.5 family has small, base and large variants for English and Chinese retrieval. The official repository documents different recommended query instructions and version changes intended to improve similarity-score behavior. A smaller model name does not directly establish higher speed on every deployment.

03

BGE rerankers are not embedding models

BAAI's reranker docs describe cross-encoder models that directly score query-document pairs, unlike an encoder that emits reusable corpus vectors. BGE-reranker-v2-m3 is multilingual; earlier base and large rerankers are separately indexed. Evaluations should measure ranking improvement, speed and how many passages are rescored.

04

Licensing and RAG deployment choices

BGE repositories in our checked index include MIT and Apache-2.0 license labels. Verify the exact model and revision, and also inspect any vector database, reranking service and inference container terms. Different retrieval representations require different indexing, storage and scoring choices.

05

Evaluating multilingual search without invented scores

BAAI publishes research and evaluation scripts for multilingual retrieval, but publisher-reported benchmark numbers are not independent OMW measurements. Reproduce the target workload, languages, hard negatives, index construction, reranker thresholds and recall/latency requirements before selecting a checkpoint.

Publisher background is attributed to official sources. Model counts and license classifications come from OMW repository evidence; no model benchmark or processing location is inferred.

MODEL-BY-MODEL REFERENCE

Which BGE model is intended for retrieval or reranking?

Browse by published modality and licensing evidence. These are navigation cues, not quality, speed or suitability rankings.

Text embeddingsmit

bge-m3

Multilingual hybrid retrieval

Inspect BGE-M3 for dense, lexical sparse and multi-vector output, and its documented preprocessing.

Commercial use: allowedLicense evidence →
Embeddings / featuresmit

bge-small-en-v1.5

Compact English embeddings

Use BGE-small-en-v1.5 as an English vector encoder and evaluate recall for your corpus.

Commercial use: allowedLicense evidence →
Embeddings / featuresmit

bge-large-en-v1.5

Larger English embeddings

Compare its vector representation, index footprint and retrieval quality with smaller BGE encoders.

Commercial use: allowedLicense evidence →
Embeddings / featuresmit

bge-small-zh-v1.5

Chinese embeddings

Review the language-specific query instructions and appropriate Chinese evaluation data.

Commercial use: allowedLicense evidence →
Text Classificationapache-2.0

bge-reranker-v2-m3

Multilingual second-stage reranker

Score candidate passages after retrieval instead of treating this as a reusable embedding encoder.

Commercial use: allowedLicense evidence →
Text Classificationmit

bge-reranker-base

Query-passage relevance scoring

Inspect this earlier cross-encoder reranker and its own license and input length.

Commercial use: allowedLicense evidence →

FIELD-BY-FIELD COMPARISON · PRIMARY LINKS

Model evidence at a glance.

Compare the exact entries in our registry. Parameter values are taken from the source-checked record; a license label is not a legal opinion, and recognized weight files are not an inference-speed benchmark. Open each record for the underlying citations and observed history.

Model and parameter countModalityLicense evidenceFormats and filesSource and history
bge-small-en-v1.50.0333605B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-m3Pending verification published parametersText embeddingsmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch4 filesRepository ↗Model evidence →History →
bge-reranker-v2-m30.567756B published parametersText Classificationapache-2.0allowed commercial-use classificationInspect license evidence →Safetensors1 filesRepository ↗Model evidence →History →
bge-base-en-v1.50.109483B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-large-en-v1.50.335142B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-small-zh-v1.50.0239544B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →PyTorch, Safetensors2 filesRepository ↗Model evidence →History →
bge-base-en0.109483B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-reranker-base0.278045B published parametersText Classificationmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-reranker-large0.559892B published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →ONNX, PyTorch, Safetensors3 filesRepository ↗Model evidence →History →
bge-large-zh-v1.5Pending verification published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →PyTorch1 filesRepository ↗Model evidence →History →
bge-base-zh-v1.5Pending verification published parametersEmbeddings / featuresmitallowed commercial-use classificationInspect license evidence →PyTorch1 filesRepository ↗Model evidence →History →

LICENSE & COVERAGE · NOT LEGAL ADVICE

How do the recorded licenses differ?

Different models from the same developer can use different license terms. The distribution below describes the labels and field-level classifications observed in the listed repositories; it does not automatically establish permissions for a specific deployment.

License labels

  • mit10 of 11
  • apache-2.01 of 11

Commercial-use classifications

  • allowed11 of 11

Model modalities

  • Embeddings / features8 of 11
  • Text Classification2 of 11
  • Text embeddings1 of 11

A classification of unknown is not a prohibition. Even a recognized permissive license can carry conditions. Review the license for the exact model and revision. Open the license explorer →

EVIDENCE METHOD · TRANSPARENT LIMITS

How this developer profile is assembled.

We aggregate the current source-checked model records, rather than publish an unverified corporate biography. Repository metadata, model cards, license files and recognized weight artifacts supply the facts where available. The record-check date does not assert that a model or its license changed on that date.

Source-derived runtime signals are not hardware benchmarks. Weight-only memory estimates omit runtime overhead, activations and other deployment requirements. Model-specific uncertainties remain visible in the individual evidence records.

QUESTIONS · DIRECT ANSWERS

Frequently asked questions about BAAI models.

Which BAAI models are covered?

Open Model Weights currently lists 11 field-verified records for BAAI. Their recorded model modalities are: 8 Embeddings / features; 2 Text Classification; 1 Text embeddings. This is a registry sample, not an exhaustive publisher catalog.

Which licenses are recorded for BAAI models?

In the 11 indexed records, repository license labels are: 10 mit; 1 apache-2.0. License labels do not replace the model-specific terms or establish rights for every use case.

Where can I verify BAAI model information?

Each indexed model links to its official source repository and a field-level Open Model Weights profile. The most recent record checks shown here reach 2026-11-13; checking a repository does not imply that its contents changed.

What does BGE-M3 support?

BAAI documents dense, lexical sparse and multi-vector retrieval representations, with multilingual inputs. The model card and inference documentation define which outputs and scoring methods are available.

Can a BGE reranker replace an embedding model in a vector database?

Not directly. BAAI's reranker expects a query-document pair to score, whereas embeddings create reusable vectors for large-scale initial retrieval. Common architectures combine both stages.

Are BAAI BGE models all licensed the same way?

No blanket software or model-license claim is safe. The checked registry includes MIT and Apache-2.0 labels on different BGE records. Consult each checkpoint's source license.

COMPLETE INDEXED SELECTION

All 11 indexed model profiles.