Text Embeddings for Semantic Search

Training encoders that turn text in many languages into vectors so that search, clustering and retrieval-augmented generation match meaning rather than keywords.

Tier Grades

Tier A: Frontier AI labs

feasible

Alibaba and BAAI train embedders of 0.6B to 8B on LLM backbones with LLM-synthesised data (Qwen3 Embedding, C-Pack).

Tier B: Indonesian AI labs

feasible

A 0.6B embedder fine-tuned from an open backbone, as in Qwen3 Embedding, is within reach for Indonesian retrieval data; Indonesian lab capacity itself is not in the sources.

Tier C: Institution units such as ITB

feasible

C-Pack shows a 6-author team building a benchmark of 35 datasets plus training data, a scale an institution unit can match for Indonesian.

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

Open Apache 2.0 embedders (Qwen3 Embedding) can be used as they are in search and RAG proofs of concept.

Where this comes from

Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

ComputeQwen3 Embedding: embedding models of 0.6B, 4B and 8B parameters built on Qwen3 LLM backbones; training compute is not given in the abstract.
DataQwen3 Embedding: multi-stage training of large-scale unsupervised pretraining then supervised fine-tuning, with Qwen3 LLMs synthesising training data across domains and languages. C-Pack: C-MTP training set and the C-MTEB benchmark of 6 tasks and 35 datasets.
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 12 authors on the Qwen3 Embedding paper; 6 on C-Pack.
Time horizonMonths.
Shelf-life signalsMTEB leadership moving from BERT-size encoders to LLM-backbone embedders (Qwen3 Embedding reports top multilingual MTEB results), Adjustable vector sizes (32 to 4,096 dimensions on Qwen3 Embedding) cutting index storage
Pipeline roleTier A releases multilingual embedders and benchmarks; Tier B and Tier C fine-tune small ones on Indonesian and domain text and build local test sets; Tier D puts them into search and RAG proofs of value.

Where this comes from

Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars

Qwen3-Embedding-8B BAAI

bge-m3

gte-multilingual-base Microsoft AI

multilingual-e5-large-instruct AI Singapore

SEA-LION-E5-Embedding-600M

Observed actors

Alibaba (Qwen) BAAI

Sentence Transformers Cohere

Cohere

Linked models

SEA-LION-E5-Embedding-600Membeddingexemplar

Qwen3-Embedding-8Bembeddingexemplar

gte-multilingual-baseembeddingexemplar

bge-m3embeddingexemplar

multilingual-e5-large-instructembeddingexemplar

EmbeddingGemma 2embeddingproduced by

SEA-LION ModernBERT-Embedding 300Membeddingproduced by

SEA-LION ModernBERT-Embedding 600Membeddingproduced by

Nemotron-3-Embed-1Bembeddingproduced by

Nemotron-3-Embed-8Bembeddingproduced by

Granite Embedding Multilingual R2embeddingproduced by

jina-embeddings-v5-omni-nanoembeddingproduced by

jina-embeddings-v5-omni-smallembeddingproduced by

Qwen3-VL-Embedding-2Bembeddingproduced by

Qwen3-VL-Embedding-8Bembeddingproduced by

BGE-VL familyembeddingproduced by

harrier-oss-v1-0.6bembeddingproduced by

jina-embeddings-v5-text-nanoembeddingproduced by

jina-embeddings-v5-text-smallembeddingproduced by

ColEmbed V2embeddingproduced by

Llama Nemotron Embed VL 1B v2embeddingproduced by

Llama Nemotron Embed 1B v2embeddingproduced by

bge-reasoner-embed-qwen3-8bembeddingproduced by

EmbeddingGemma 300Membeddingproduced by

+72

Granite Embedding small English R2embeddingproduced by

Qwen3-Embedding-0.6Bembeddingproduced by

Qwen3-Embedding-4Bembeddingproduced by

jina-embeddings-v4embeddingproduced by

Gemini Embeddingembeddingproduced by

gemini-embedding-001embeddingproduced by

nomic-embed-multimodalembeddingproduced by

Nomic embed text v2 MoEembeddingproduced by

Granite Embeddingembeddingproduced by

Granite Embedding 278M multilingualembeddingproduced by

ColQwen2 v1.0embeddingproduced by

jina-embeddings-v3embeddingproduced by

NV-Embed-v2embeddingproduced by

ColPali v1.2embeddingproduced by

ColPaliembeddingproduced by

gte-Qwen2-7B-instructembeddingproduced by

gte-large-en-v1.5embeddingproduced by

snowflake-arctic-embed-lembeddingproduced by

Snowflake Arctic Embedembeddingproduced by

snowflake-arctic-embed-membeddingproduced by

mxbai-embed-large-v1embeddingproduced by

nomic-embed-text-v1.5embeddingproduced by

nomic-embed-text-v1embeddingproduced by

Nomic Embed Text v1embeddingproduced by

SFR-Embedding-Mistralembeddingproduced by

e5-mistral-7bembeddingproduced by

jina-embeddings-v2-base-enembeddingproduced by

bge-large-en-v1.5embeddingproduced by

bge-large-zh-v1.5embeddingproduced by

bge-small-en-v1.5embeddingproduced by

bge-base-en-v1.5embeddingproduced by

BGE v1.5embeddingproduced by

GTE (General Text Embeddings)embeddingproduced by

gte-largeembeddingproduced by

gte-smallembeddingproduced by

multilingual-e5-smallembeddingproduced by

all-MiniLM-L6-v2 (Xenova, Transformers.js ONNX)embeddingproduced by

SPECTER2embeddingproduced by

SPLADE (CoCondenser self-distil)embeddingproduced by

paraphrase-multilingual-MiniLM-L12-v2embeddingproduced by

paraphrase-multilingual-mpnet-base-v2embeddingproduced by

GTR (Generalizable T5 Retriever)embeddingproduced by

Sentence-T5embeddingproduced by

SapBERTembeddingproduced by

LaBSEembeddingproduced by

LASERembeddingproduced by

Universal Sentence Encoderembeddingproduced by

fastText word vectorsembeddingproduced by

GloVeembeddingproduced by

bge-multilingual-gemma2embeddingproduced by

E5embeddingproduced by

Mistral Embedembeddingproduced by

Nemotron Retriever modelsembeddingproduced by

Amazon Titan Text Embeddings V2embeddingproduced by

Cohere Embedembeddingproduced by

Embed English v3.0embeddingproduced by

Embed Multilingual v3.0embeddingproduced by

Embed v4.0embeddingproduced by

Embed v5.0 Proembeddingproduced by

Gemini Embedding 2embeddingproduced by

OpenAI text embeddingsembeddingproduced by

text-embedding-3-largeembeddingproduced by

text-embedding-3-smallembeddingproduced by

text-embedding-ada-002embeddingproduced by

all-MiniLM-L12-v2embeddingproduced by

all-MiniLM-L6-v2embeddingproduced by

all-mpnet-base-v2embeddingproduced by

voyage-3.5embeddingproduced by

voyage-3-largeembeddingproduced by

voyage-4embeddingproduced by

voyage-code-3embeddingproduced by

Voyage embeddingsembeddingproduced by

Papers

Sentence-BERT: Sentence Embeddings using Siamese BERT-NetworksNils Reimers et al., EMNLP 2019 (per arXiv comment), 2019 founding
Dense Passage Retrieval for Open-Domain Question AnsweringVladimir Karpukhin et al., EMNLP 2020 (per arXiv comment), 2020 founding
Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation ModelsYanzhao Zhang et al., arXiv preprint, 2025 state of the art
C-Pack: Packed Resources For General Chinese EmbeddingsShitao Xiao et al., SIGIR 2024 (per arXiv comment), 2023 state of the art
When Text Embedding Meets Large Language Model: A Comprehensive SurveyZhijie Nie et al., arXiv preprint, 2024 survey

Where this comes from

The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

arxiv.org/abs/2506.05176 https://arxiv.org/abs/2506.05176
arxiv.org/abs/2309.07597 https://arxiv.org/abs/2309.07597

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Text Embeddings for Semantic Search", https://aurion.semai.id/topics/text-embeddings-for-semantic-search/, records of October 2026.

Text Embeddings for Semantic Search

Tier Grades

Tier A: Frontier AI labs

Tier B: Indonesian AI labs

Tier C: Institution units such as ITB

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

Anchors

Who and What

Links

Models 96

Papers

Sources