bge-m3

BAAI MIT out 3 years ago

Text in, embeddings out, 8K tokens of context. 3 providers serve it from $0.010 per million input tokens.

Verified 2026-10-25 Rank #2 in Embeddings · score 123.1 method v1 Sources 4 Downloads 30d 29.8M · #3 of 619 tracked

What does it cost?$0.010/1M incheapest of 1DeepInfrafp322026-10-24

Can I use it?MITPermissivelicence file2026-10-24

Does it work?63.21on MTEB#20 of 64 measuredleaderboard2026-10-24

Do people use it?29.8Mdownloads, 30 days#3 of 619 we track · top 1%+23 likes, downloads −7% this week3,925 likesHugging Face2026-10-24

Worth reading

Picked by rule from 12 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Hard-coding BERT token IDs into a BGE-M3 ONNX pipeline causes the model to receive invalid inputs.

From Understanding BGE-M3 Embedding Inference and ONNX Runtime Tuning | wiphoo.dev · wiphoo.dev · 2026-06-04

Latest · 5 months ago

Understanding BGE-M3 Embedding Inference and ONNX Runtime Tuning | wiphoo.dev

BGE-M3 generates dense embeddings with a hidden_size of 1024 by converting text to tokens and running inference via ONNX Runtime.

Hard-coding BERT token IDs into a BGE-M3 ONNX pipeline causes the model to receive invalid inputs.
The first few inference runs are slower because the memory arena has not yet reached its working size.

wiphoo.dev 2026-06-04 2.8k words

12 best of 12 tracked

The 2 picks above, plus these 10. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleBGE-M3 vs OpenAI Embeddings [2026]: Free & Multilingual · Aidxnaidxn.com · 2026-05-12
ArticleBGE-M3 and BGE Reranker in 2026: what the benchmarks say about dense, lexical, and multi-vector retrievalaxiomlogica.com · 2026-05-05
ArticleBGE-M3: Unified Multilingual Retrieval Modelemergentmind.com · undated
Articlebge-m3 by BAAI - Overview, Benchmarks & Hardware | Q4KM.aiq4km.ai · undated
ArticleBGE M3 — Specs, Pricing & GPU Requirements | InferenceBench | InferenceBenchinferencebench.io · undated
ArticleQuantized bge-M3 embedding model in the browser for full privacyciekawy.github.io · 2024-07-21 · 2 HN pts
ArticleBGE-M3 Embeddings: Dense, Sparse & Multi-Vector in 2026pristren.com · 2026-03-13
ArticleUnleash the power of sparse vector: Hybrid search using pgvecto.rs and BGE-M3blog.pgvecto.rs · 2024-07-11
ArticleBGE-M3 vs. Spladezilliz.com · 2024-04-29 · 1 HN pts
ArticleBGE-M3 — what it needs and what runs it · AITrackerbluetechgreen.com · 2024-01-30

Show all 10

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#	Cheapest 3	In /1M	Out /1M	Context	Quant	Seen
1	DeepInfra	$0.010	—	8K	fp32	Oct 24
2	HF Inference	—	—	—	—	Oct 25
3	Parasail	—	—	—	—	Oct 24

All 3 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 3 API providers; heavily run locally (303 quantized builds); a big base for builders (617 fine-tunes, adapters and merges).

As a service3DeepInfra, HF Inference, Parasail
On own hardware303quantized builds
Built on617582 fine-tunes · 34 adapters · 1 merges
In apps100+Hugging Face Spaces

Top Spaces running it: embedl/hfviewer ♥ 181 · SAkizuki/DanbooruSearch ♥ 179 · SAkizuki/WenQiuYue ♥ 27 · antoinelouis/mtem-pruner ♥ 14

How it scores

Third-party measurements only; author-reported numbers stay in Reference, marked as such.

#20 of 64 on MTEB so far

63.21. Each tick is a model on this site with a measured MTEB row; author-reported numbers stay in Reference. Mean over 27/41 tasks present. 27/41 tasks

Source · leaderboard · 2026-10-24

The story so far

Weights, every dated article, and repo commits on one rail.

6 dated articles and 30 repo commits so far

Teal is the weights (2024-01-27), amber triangles are 6 dated articles (each links to it), grey ticks are repo commits.

What people built on it

920 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

920 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF ggml-org/bge-m3-Q8_0-GGUF for llama.cpp, Ollama, LM Studio 111.1k downloads MLX mlx-community/bge-m3-mlx-fp16 for Apple silicon 42.9k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

1 row

Benchmarks 1 row

measured third-party leaderboard row (1)author-reported number from the model card or eval filesNo confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
measured MTEB · Embed text for search	63.21 mean over 27/41 tasks present	27/41 tasks	leaderboard	2026-10-24

What the card says

2 cited sentences

Summary generated, cited

The model supports more than 100 working languages and processes inputs up to 8192 tokens.^[1]^[2] It is released under the MIT license and can perform dense, multi-vector, and sparse retrieval.^[3]^[4]

“support more than 100 working languages”
“up to 8192 tokens”
license
“dense retrieval, multi-vector retrieval, and sparse retri…”

Generated 2026-10-07 14:59 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

2 use cases

What it is good at 2 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

18 fields

Listed in Embeddings

Facts

Context length is 8,194 tokens (8K). (config.json max_position_embeddings, observed 2026-10-24)
License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
Created 2024-01-27, last modified 2024-07-03. (HF API, observed 2026-10-24)

Pipeline tag	sentence-similarity
Library	sentence-transformers
Architectures	XLMRobertaModel
Model type	xlm-roberta
Dtype	float32
MoE	no
Context length	8,194 (config.json max_position_embeddings)
Papers	arXiv:2402.03216, arXiv:2004.04906, arXiv:2106.14807, arXiv:2107.05720, arXiv:2004.12832
License	MIT (mit)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · XLMRobertaModel · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · XLMRobertaModel · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

All variants

40 variants

Lineage

Original model Family: BAAI/bge-m3 920 derivatives on the Hub

finetune: 582quantized: 303adapter: 34merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
upskyy/bge-m3-korean	finetune	—	568M	S	—	upskyy	66	10.3M
littlejohn-ai/bge-m3-spa-law-qa	finetune	—	568M	S	apache-2.0	littlejohn-ai	22	3.7M
dragonkue/BGE-m3-ko	finetune	—	568M	S	apache-2.0	dragonkue	79	3.6M
nlpai-lab/KURE-v1	finetune	—	568M	S	mit	nlpai-lab	100	3.1M
AITeamVN/Vietnamese_Embedding	quantized	—	568M	S	apache-2.0	AITeamVN	70	1.5M
Xenova/bge-m3	quantized	—	—	—	mit	Xenova	53	1.1M
dariolopez/bge-m3-es-legal-tmp-6	finetune	—	568M	S	apache-2.0	dariolopez	2	517.8k
antoinelouis/french-bge-m3	quantized	—	350M	S	mit	antoinelouis	0	432.1k
namdp-ptit/ViRanker	finetune	—	568M	S	apache-2.0	namdp-ptit	17	346.1k
farbodtavakkoli/OTel-Embedding-568M	finetune	—	568M	S	apache-2.0	farbodtavakkoli	0	335.1k
AITeamVN/Vietnamese_Embedding_v2	quantized	—	568M	S	apache-2.0	AITeamVN	13	308.9k
ggml-org/bge-m3-Q8_0-GGUF	quantized	Q8_0	—	—	mit	ggml-org	21	111.1k
OpenMed/OpenMed-PII-Italian-ClinicalBGE-Large-568M-v1	finetune	—	567M	S	apache-2.0	OpenMed	0	96.2k
mhaseeb1604/bge-m3-law	finetune	—	568M	S	—	mhaseeb1604	6	94.6k
sayed0am/arabic-english-bge-m3	quantized	—	362M	S	mit	sayed0am	9	77.9k
OpenMed/OpenMed-PII-Spanish-ClinicalBGE-Large-568M-v1	finetune	—	567M	S	apache-2.0	OpenMed	0	70.6k
gpahal/bge-m3-onnx-int8	quantized	int8	—	—	mit	gpahal	3	47.7k
mlx-community/bge-m3-mlx-fp16	finetune	mlx	568M	S	mit	mlx-community	4	42.9k
OpenMed/OpenMed-PII-German-ClinicalBGE-Large-568M-v1	finetune	—	567M	S	apache-2.0	OpenMed	0	41.6k
ekacare/parrotlet-e	finetune	—	568M	S	cc-by-sa-4.0	ekacare	1	41k
OpenMed/OpenMed-PII-French-ClinicalBGE-Large-568M-v1	finetune	—	567M	S	apache-2.0	OpenMed	0	40.6k
NohTow/french-bge-m3	quantized	—	358M	S	mit	NohTow	1	27.4k
BAAI/CCI3-HQ-Classifier	finetune	—	568M	S	apache-2.0	BAAI	13	23.8k
OpenMed/OpenMed-PII-ClinicalBGE-Large-568M-v1	finetune	—	567M	S	apache-2.0	OpenMed	0	23.1k
groonga/bge-m3-Q4_K_M-GGUF	quantized	GGUF	—	—	mit	groonga	2	19.7k
nezahatkorkmaz/turkce-embedding-bge-m3	finetune	—	568M	S	—	nezahatkorkmaz	11	16.9k
Roflmax/bge-m3-legal-ru-updata	finetune	—	568M	S	—	Roflmax	0	16.2k
yaniseuranova/setfit-rag-hybrid-search-query-router	finetune	—	568M	S	—	yaniseuranova	2	16k
nonola/portuguese-bge-m3	quantized	—	349M	S	mit	nonola	2	14.6k
KimChen/bge-m3-GGUF	quantized	GGUF	—	—	mit	KimChen	18	13.6k
hamtaai/bg3_model	finetune	—	568M	S	—	hamtaai	0	14.4k
mlx-community/bge-m3-mlx-8bit	finetune	mlx	568M	S	mit	mlx-community	2	12.5k
fineinstructions/instruction_template_retrieval_embedding	finetune	—	568M	S	—	fineinstructions	0	12.5k
cstr/bge-m3-GGUF	quantized	GGUF	—	—	mit	cstr	1	10.1k
altaidevorg/bge-m3-distill-8l	finetune	—	366M	S	—	altaidevorg	12	9.1k
onnx-community/bge-m3-ONNX	quantized	—	—	—	mit	onnx-community	3	8.8k
MahradHosseini/bge-m3-onnx-int8	quantized	int8	—	—	apache-2.0	MahradHosseini	2	7.9k
philipchung/bge-m3-onnx	quantized	—	—	—	mit	philipchung	3	7.7k
bbvch-ai/bge-m3-GGUF	quantized	GGUF	—	—	mit	bbvch-ai	4	7.3k
gbnath/english-bge-m3	quantized	—	353M	S	mit	gbnath	2	6.1k

Repository changelog

30 commits

Repository changelog 30 commits

Date	Commit	SHA
2024-07-03	Update MIRACL evaluation results of BGE-M3 (#67)	5617a9f
2024-07-02	Update MIRACL evaluation results (#68)	e1f4568
2024-04-13	Adding ONNX file of this model (#44)	babcf60
2024-04-02	Delete model.safetensors	c400e3d
2024-04-02	Delete bge-m3-with-deocder	7b44cab
2024-04-02	Upload bge-m3-with-deocder/pytorch_model.bin with huggingface_hub	80087bf
2024-03-26	Update tokenizer_config.json	50f9396
2024-03-20	Update README.md	5a21248
2024-03-19	Update README.md	cab5a40
2024-03-08	Upload others.webp	ecbfce3
2024-03-08	Update README.md	2ab44ec
2024-03-02	Update README.md	3641912
2024-02-11	Update README.md	3ab7155
2024-02-11	Upload bm25.jpg	f2dfce0
2024-02-11	Update README.md	cfdb103
2024-02-08	Upload 5 files	aa47896
2024-02-08	Update README.md	6d44202
2024-02-07	Update README.md	2d5552f
2024-02-06	Upload long.jpg	3c54f5a
2024-02-06	Upload long.jpg	96e9e91
2024-02-06	Update README.md	5cedd82
2024-02-06	Update README.md	f11a3f1
2024-02-05	Update README.md	8017dbb
2024-02-05	Update README.md	d5ab3da
2024-02-01	Update README.md	4277867
2024-02-01	Update README.md	1f5d3ac
2024-02-01	Update README.md	a12bcd7
2024-02-01	Update README.md	0537228
2024-01-31	Update README.md	3806044
2024-01-31	Upload nqa.jpg	84790c1

All providers

3 providers · 3 endpoints

Where to run 3 providers · 3 endpoints

Hosted endpoints observed for bge-m3, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.01/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
DeepInfra deepinfra	BAAI/bge-m3	8K	fp32	$0.01	—	—	—	—	live	2026-10-24
HF Inference hf	BAAI/bge-m3	—	—	—	—	—	—	—	live	2026-10-25
Parasail parasail	parasail-bge-m3	—	—	—	—	—	—	—	serverless	2026-10-24

Articles and coverage

12 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis10

Understanding BGE-M3 Embedding Inference and ONNX Runtime Tuning | wiphoo.dev

wiphoo.dev2026-06-042,825 words

BGE-M3 generates dense embeddings with a hidden_size of 1024 by converting text to tokens and running inference via ONNX Runtime. The model uses CLS pooling to extract the vector from the first token position before L2 normalization. Tuning intra_op_threads controls parallelism within single operations, while a shared session model prevents thread oversubscription across concurrent requests.written, reviewed, cited

Hard-coding BERT token IDs into a BGE-M3 ONNX pipeline causes the model to receive invalid inputs.
The first few inference runs are slower because the memory arena has not yet reached its working size.
With the default ORT_SEQUENTIAL execution mode, ONNX Runtime does not create an inter-op pool at all.

BGE-M3 vs OpenAI Embeddings [2026]: Free & Multilingual · Aidxn

aidxn.com2026-05-121,043 words

BGE-M3 is an MIT licensed, open-weights model that took the top of the MTEB leaderboard and beats closed-source incumbents on most retrieval benchmarks. It outputs 1024-dim vectors and supports 100+ languages in one shared vector space for cross-language retrieval without translation.written, reviewed, cited

BGE-M3 produces dense, sparse, and ColBERT-style vectors from a single forward pass.
Hybrid retrieval with BGE-M3 requires only a flag instead of merging separate keyword and vector results.
Switching to BGE-M3 requires re-embedding the entire corpus because vectors from different models live in different spaces.

BGE-M3 and BGE Reranker in 2026: what the benchmarks say about dense, lexical, and multi-vector retrieval

axiomlogica.com2026-05-05

BGE-M3: Unified Multilingual Retrieval Model

emergentmind.com2,085 words

BGE-M3 unifies dense, sparse, and multi-vector retrieval modes in a single encoder built on XLM-RoBERTa-large with 550M parameters and a vocabulary of 250,000 subword tokens. The model extends maximum input length to 8,192 tokens, is further pre-trained via RetroMAE on 184M unlabeled samples, and uses split-batch training to preserve large in-batch negative pools. For memory-constrained deployment, it applies the MCLS trick, while downstream SafeGen fine-tuning achieves Accuracy = 0.8215 on held-out test sets.written, reviewed, cited

Disabling self-knowledge distillation and supervising each head separately hurts the sparse head by approximately 17 points in nDCG@10 on MIRACL.
In SafeGen, a fine-tuned BGE-M3 classifier achieves Accuracy = 0.8215 and F1-Score = 0.8145 versus 0.1865 and 0.1840 for the pretrained base.

BGE-M3: Multilingual Bi-Encoder Model

emergentmind.com1,336 words

BGE-M3 is a 550M parameter multilingual bi-encoder supporting dense, sparse, and multi-vector retrieval for inputs up to 8,192 tokens. The model achieves a mean RAGAS score of 70.99 on Arabic datasets and typically requires 5–10 ms inference time for 512-token queries on an A100. Engineers should index embeddings in a FAISS index using cosine similarity for efficient retrieval.written, reviewed, cited

The model architecture is derived from XLM-RoBERTa-large and enhanced with retrofitted masked autoencoding (RetroMAE).
All Arabic experiments are conducted on Modern Standard Arabic, leaving performance on dialect variants unevaluated.

bge-m3 by BAAI - Overview, Benchmarks & Hardware | Q4KM.ai

q4km.ai983 words

BGE M3 — Specs, Pricing & GPU Requirements | InferenceBench | InferenceBench

inferencebench.io575 words

Quantized bge-M3 embedding model in the browser for full privacy

ciekawy.github.io2024-07-21HN 2 pts · 1 comments160 words

BGE-M3 Embeddings: Dense, Sparse & Multi-Vector in 2026

pristren.com2026-03-13771 words

BGE-M3 vs. Splade

zilliz.com2024-04-29HN 1 pts · 0 comments1,900 words

All coverage tracked for this model (12) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
BGE-M3: Unified Multilingual Retrieval Model	emergentmind.com	—	2026-10-07	exa
BGE-M3: Multilingual Bi-Encoder Model	emergentmind.com	—	2026-10-07	exa
bge-m3 by BAAI - Overview, Benchmarks & Hardware \| Q4KM.ai	q4km.ai	—	2026-10-07	exa
BGE M3 — Specs, Pricing & GPU Requirements \| InferenceBench \| InferenceBench	inferencebench.io	—	2026-10-07	exa
Understanding BGE-M3 Embedding Inference and ONNX Runtime Tuning \| wiphoo.dev	wiphoo.dev	2026-06-04	2026-10-07	exa
BGE-M3 vs OpenAI Embeddings [2026]: Free & Multilingual · Aidxn	aidxn.com	2026-05-12	2026-10-07	exa
BGE-M3 and BGE Reranker in 2026: what the benchmarks say about dense, lexical, and multi-vector retrieval	axiomlogica.com	2026-05-05	2026-10-07	exa
BGE-M3 Embeddings: Dense, Sparse & Multi-Vector in 2026	pristren.com	2026-03-13	2026-10-07	exa
Quantized bge-M3 embedding model in the browser for full privacy	ciekawy.github.io	2024-07-21	2026-10-07	hn
Unleash the power of sparse vector: Hybrid search using pgvecto.rs and BGE-M3	blog.pgvecto.rs	2024-07-11	2026-10-07	hn
BGE-M3 vs. Splade	zilliz.com	2024-04-29	2026-10-07	hn
BGE-M3 — what it needs and what runs it · AITracker	bluetechgreen.com	2024-01-30	2026-10-07	exa

Papers

5 papers

Papers 5

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Dense Passage Retrieval for Open-Domain Question Answering cited in card peer-reviewed · measured

Vladimir Karpukhin, Barlas Oğuz, Sewon Min +5 · 2020-04-10 · Conference on Empirical Methods in Natural Language Processing (EMNLP) · PDF · HF paper page (3 upvotes) · evidence
M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation cited in card

Jianlv Chen, Shitao Xiao, Peitian Zhang +3 · 2024-02-05 · PDF · HF paper page (10 upvotes)
ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT cited in card

Omar Khattab, Matei Zaharia · 2020-04-27 · PDF · HF paper page (7 upvotes)
SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking cited in card

Thibault Formal, Benjamin Piwowarski, Stéphane Clinchant · 2021-07-12 · PDF · HF paper page (1 upvotes)
A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques cited in card

Jimmy Lin, Xueguang Ma · 2021-06-28 · PDF · HF paper page

Similar models

3 models

Similar models

bge-large-zh-v1.5
fewer providers, permissive license
codebert-base
fewer providers
bart-large
fewer providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 UTC · Last verified 2026-10-25

Follow bge-m3

Watch this model

/feeds/models/BAAI--bge-m3.xml(RSS 2.0) · every change to bge-m3
/feeds/families/BAAI--bge-m3.xml(RSS 2.0) · every change across the bge-m3 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.