The story so far
Weights, every dated article, and repo commits on one rail.
22 repo commits since the weights landed
Teal is the weights (2022-03-02), grey ticks are repo commits.
LaBSE
sentence-transformers Apache 2.0 out 5 years ago
471M parameters, text in, embeddings out, 512 tokens of context. 1 provider serves it.
Verified 2026-10-25 Rank #55 in Embeddings · score 87.9 method v1 Sources 3 Downloads 30d 627.4k · #148 of 619 tracked
Do people use it?627.4kdownloads, 30 days#148 of 619 we track · top 24%−1 likes, downloads +16% this week349 likesHugging Face2026-10-24
Where to run it
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
| # | Cheapest live row | In /1M | Out /1M | Context | Quant | Seen |
|---|---|---|---|---|---|---|
| 1 | HF Inference | — | — | — | — | Oct 25 |
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 1 API provider; 8 quantized builds for local use; a big base for builders (103 fine-tunes, adapters and merges).
- As a service1HF Inference
- On own hardware8quantized builds
- Built on103103 fine-tunes · 0 adapters · 0 merges
- In apps100+Hugging Face Spaces
Top Spaces running it: SeaLLMs/SeaLLM-Chat ♥ 38 · as-cle-bert/pokemon-bot ♥ 6 · ReithBjarkan/SEO_Keyword_Similarity_Tool ♥ 6 · kaengreg/rusBEIR ♥ 6
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 0.3 GB at INT4, 0.6 GB at FP8, 1.1 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
How it scores
Third-party measurements only; author-reported numbers stay in Reference, marked as such.
#54 of 64 on MTEB so far
48.81. Each tick is a model on this site with a measured MTEB row; author-reported numbers stay in Reference. Mean(task) x100 over 41 tasks. 41 tasks, split=test
Source · leaderboard · 2026-10-24
What people built on it
111 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
111 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.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
3 rows
Benchmarks 3 rows
measured third-party leaderboard row (3)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 | 18.05 nDCG@10 x100 (mean of 10 tasks) | 10 tasks, split=test | leaderboard | 2026-10-24 |
| measured MTEB · Embed text for search | 48.81 mean(task) x100 over 41 tasks | 41 tasks, split=test | leaderboard | 2026-10-24 |
| measured MTEB Reranking · Rerank search results | 41.28 MAP x100 (mean of 2 tasks) | 2 tasks, split=test | leaderboard | 2026-10-24 |
Use-case evidence
3 use cases
What it is good at 3 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
20 fields
Listed in Embeddings
Facts
- LaBSE has 471M parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 512 tokens (512). (config.json max_position_embeddings, observed 2026-10-24)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
- Created 2022-03-02, last modified 2025-03-06. (HF API, observed 2026-10-24)
| Pipeline tag | sentence-similarity |
|---|---|
| Library | sentence-transformers |
| Architectures | BertModel |
| Model type | bert |
| Params by dtype | F32: 470,926,848 · I64: 512 |
| Size band | S |
| MoE | no |
| Context length | 512 (config.json max_position_embeddings) |
| Languages | multilingual, af, sq, am, ar, hy, as, az, eu, be, bn, bs, bg, my, ca, ceb, zh, co, hr, cs, da, nl, en, eo, et, fi, fr, fy, gl, ka, de, el, gu, ht, ha, haw, he, hi, hmn, hu, is, ig, id, ga, it, ja, jv, kn, kk, km, rw, ko, ku, ky, lo, la, lv, lt, lb, mk, mg, ms, ml, mt, mi, mr, mn, ne, false, ny, or, fa, pl, pt, pa, ro, ru, sm, gd, sr, st, sn, si, sk, sl, so, es, su, sw, sv, tl, tg, ta, tt, te, th, bo, tr, tk, ug, uk, ur, uz, vi, cy, wo, xh, yi, yo, zu |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · BertModel · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · BertModel · 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.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for LaBSE: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 1.1 GB · Edge (<4 GB) |
|---|---|
| FP8 (1 byte/param) | 0.6 GB · Edge (<4 GB) |
| INT4 (0.5 bytes/param) | 0.3 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
40 variants
Lineage
quantized: 8finetune: 103
Top variants (40 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| minishlab/M2V_multilingual_output | quantized | — | 128M | S | mit | minishlab | 19 | 335.6k |
| Xenova/LaBSE | quantized | — | — | — | — | Xenova | 2 | 23.8k |
| Omartificial-Intelligence-Space/Arabic-labse-Matryoshka | finetune | — | 471M | S | apache-2.0 | Omartificial-Intelligence-Space | 5 | 16.1k |
| data-silence/any-news-classifier | finetune | — | 471M | S | apache-2.0 | data-silence | 0 | 6k |
| fredxlpy/LuxEmbedder | finetune | — | 471M | S | cc-by-nc-4.0 | fredxlpy | 2 | 3.2k |
| LocalDoc/TEmA-small | finetune | — | — | — | cc-by-4.0 | LocalDoc | 0 | 3.2k |
| data-silence/frozen_news_classifier_ft | finetune | — | 471M | S | apache-2.0 | data-silence | 0 | 2.5k |
| artefucktor/LaBSE_geonames_RU | finetune | — | 471M | S | — | artefucktor | 1 | 925 |
| MWirelabs/ne-embed | finetune | — | 471M | S | cc-by-4.0 | MWirelabs | 1 | 818 |
| lemon-mint/LaBSE-EnKo-Nano-Preview-v0.3 | finetune | — | 126M | S | apache-2.0 | lemon-mint | 3 | 716 |
| artefucktor/LaBSE_geonames_RU_RELOCATION | finetune | — | 471M | S | apache-2.0 | artefucktor | 0 | 766 |
| lingtrain/labse-udmurt | finetune | — | 471M | S | — | lingtrain | 4 | 540 |
| LocalDoc/LaBSE-small-AZ | finetune | — | 142M | S | apache-2.0 | LocalDoc | 0 | 695 |
| dima-does-code/LaBSE-geonames-15K-MBML-5e-v1 | finetune | — | 471M | S | apache-2.0 | dima-does-code | 0 | 690 |
| Singaraj/morisien-embed-v1.5 | quantized | — | 471M | S | mit | Singaraj | 3 | 440 |
| NetherQuartz/LaBSE-tokipona | finetune | — | 471M | S | — | NetherQuartz | 3 | 399 |
| midwestcyr/labse-ruskaromani | finetune | — | 471M | S | mit | midwestcyr | 1 | 437 |
| leolabat/opibert-scl-cls | finetune | — | 471M | S | — | leolabat | 0 | 448 |
| bicolino34/LaBSE-ja-uk | finetune | — | 471M | S | — | bicolino34 | 1 | 387 |
| lingtrain/labse-chuvash | finetune | — | 471M | S | — | lingtrain | 0 | 429 |
| sanganaka/LaBSE-sanskritFT | finetune | — | 471M | S | — | sanganaka | 0 | 411 |
| dima-does-code/LaBSE-geonames-15K-MBML-1e-v1 | finetune | — | 471M | S | apache-2.0 | dima-does-code | 0 | 390 |
| ahbond/lebse | finetune | — | 471M | S | apache-2.0 | ahbond | 1 | 333 |
| Nimsara2001/labse-sinhala-finetuned | finetune | — | 471M | S | — | Nimsara2001 | 1 | 332 |
| codersan/FaLaBSE-v3 | finetune | — | 471M | S | — | codersan | 0 | 381 |
| codersan/FaLaBSE-v12-phase1-Quora | finetune | — | 471M | S | — | codersan | 0 | 380 |
| lingtrain/labse-sakha | finetune | — | 471M | S | — | lingtrain | 1 | 322 |
| Sammau/labse-dhivehi-finetuned | finetune | — | 471M | S | — | Sammau | 0 | 369 |
| HSE-Chukchi-NLP/LaBSE-v2-russian-chukchi | finetune | — | 471M | S | — | HSE-Chukchi-NLP | 0 | 363 |
| adeshkin/labse-kjh-ru-mnrl-1 | finetune | — | 471M | S | — | adeshkin | 0 | 361 |
| panagoa/LaBSE-kbd-v0.2 | finetune | — | 471M | S | — | panagoa | 0 | 359 |
| dima-does-code/LaBSE-geonames-15K-MBML-5e-v2 | finetune | — | 471M | S | apache-2.0 | dima-does-code | 0 | 355 |
| codersan/FaLaBSE-v6 | finetune | — | 471M | S | — | codersan | 0 | 354 |
| Adel-Elgabry/arabic-semantic-search | finetune | — | 125M | S | — | Adel-Elgabry | 0 | 351 |
| codersan/validadted_FaLabse_onV8d | finetune | — | 471M | S | — | codersan | 0 | 346 |
| shiimi/labse-dhivehi-finetuned | finetune | — | 471M | S | — | shiimi | 0 | 343 |
| dima-does-code/LaBSE-geonames-15K-MBML-10e-v1 | finetune | — | 471M | S | apache-2.0 | dima-does-code | 0 | 340 |
| CMU-Wav2Gloss/Gitksan-encoder-bsz128-e1k-bsz32-e1k | finetune | — | 471M | S | — | CMU-Wav2Gloss | 0 | 337 |
| codersan/FaLaBSE-v1 | finetune | — | 471M | S | — | codersan | 0 | 333 |
| Jarbas/m2v-256-LaBSE | finetune | — | 128M | S | mit | Jarbas | 0 | 331 |
Repository changelog
22 commits
Repository changelog 22 commits
| Date | Commit | SHA |
|---|---|---|
| 2025-03-06 | Remove deprecated (SEB) evaluation results section | 836121a |
| 2024-03-27 | Adding `safetensors` variant of this model (#6) | b7f9471 |
| 2024-10-17 | Add new SentenceTransformer model with an onnx backend | ac91d67 |
| 2024-06-03 | Revert "Adding `safetensors` variant of this model (#6)" | e34fab6 |
| 2024-03-27 | Adding `safetensors` variant of this model (#6) | 50fe094 |
| 2024-02-25 | Changed "labse-pytorch/special_tokens_map.json" to "special_tokens_map.json" since there is no "labse-pytorch" directory. (#5) | 5513ed8 |
| 2024-02-12 | Update metadata manually | 3ab788e |
| 2023-11-02 | Update README.md (#3) | 4f2f77c |
| 2022-06-15 | Add TF weights | 931b5f9 |
| 2021-08-05 | Update README.md | 5245bbd |
| 2021-08-05 | Update README.md | e80c3a5 |
| 2021-06-22 | Add new SentenceTransformer model. | 61966e1 |
| 2021-06-03 | Update README.md | f84a947 |
| 2021-05-20 | upload flax model | 5a4feb1 |
| 2021-05-20 | allow flax | a315492 |
| 2020-12-11 | Migrate model card from transformers-repo | 4572cc2 |
| 2020-10-12 | Update vocab.txt | 867721c |
| 2020-10-12 | Update special_tokens_map.json | 0c5b145 |
| 2020-10-12 | Update tokenizer_config.json | f0eff48 |
| 2020-10-12 | Update config.json | 9a3922d |
| 2020-10-12 | Update pytorch_model.bin | c3d93bf |
| 2020-10-12 | initial commit | 152c68a |
All providers
1 provider · 1 endpoint
Where to run 1 providers · 1 endpoints
Hosted endpoints observed for LaBSE, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| HF Inference hf | sentence-transformers/LaBSE | — | — | — | — | — | — | — | live | 2026-10-25 |
Articles and coverage
1 piece
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.
Analysis1
emergentmind.com1,328 words
LaBSE is a multilingual dual-encoder model that produces language-independent sentence embeddings for 109+ languages using a shared BERT-based architecture. It offers 12-layer base or 24-layer large variants and achieves an 80% reduction in required parallel training data compared to models without such pretraining. The model supports practical applications such as parallel data mining, semantic search, and fine-grained word alignment for effective multilingual NLP.written, reviewed, cited
- LaBSE uses a shared vocabulary of either 119,547 or 501,153 tokens to support over 109 languages.
- LaBSE attains peak UN performance with ~200M bitext examples versus ~1B for models without such pretraining.
- LaBSE achieves an average Alignment Error Rate of 18.8% outperforming mBERT (22.3%) and XLM-R (27.9%).
Similar models
3 models
Similar models
- gte-large-en-v1.5 · 434M same size band (S), same task (sentence-similarity), permissive license
- snowflake-arctic-embed-l-v2.0 · 568M same size band (S), same task (sentence-similarity), permissive license
- e5-large-v2 · 335M same size band (S), same task (sentence-similarity), more providers, permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:03 UTC · Last verified 2026-10-25
Follow LaBSE
Watch this model
- /feeds/models/sentence-transformers--LaBSE.xml(RSS 2.0) · every change to LaBSE
- /feeds/families/sentence-transformers--LaBSE.xml(RSS 2.0) · every change across the LaBSE family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.