Best Embedding Models in 2026 on the Live MTEB Leaderboard

Best Embedding Models in October 2026

On the live MTEB English v2 leaderboard (checked October 14, 2026), the top open-weight general embedding models include Qwen3-Embedding-8B (Apache 2.0, mean task score 75.23) and QZhou-Embedding, while the top multilingual v2 model is Microsoft’s harrier-oss-v1-27b (MIT, 74.27). For a RAG team the right pick depends on license, language and hosting, so test two or three on your own queries.

Why embedding models are the foundation of modern AI

Embedding models translate text (or images, or audio) into dense numerical vectors that capture semantic meaning. If embeddings are new to you, start with what embeddings are. Three workloads depend on them:

1. Retrieval-augmented generation (RAG): every chunk in your knowledge base needs to be embedded; every query needs to be embedded; cosine similarity surfaces the relevant chunks.
2. Semantic search: traditional keyword search misses synonyms, paraphrases, and intent; embeddings catch them.
3. Clustering and classification: customer-support tickets, product catalog, news articles, anything that needs to be grouped by meaning.

Types of embedding models: 2026 view

Static word embeddings (Word2Vec, GloVe, FastText)

• Word2Vec: Google’s 2013 neural-net approach to word vectors.
• GloVe: Stanford’s global matrix factorisation plus local context windows.
• FastText: Facebook’s subword-aware extension of Word2Vec.

Verdict in 2026: static word embeddings are legacy. They give every occurrence of “bank” the same vector regardless of whether the context is financial or fluvial. Use them only for historical reproducibility, not new builds.

Contextual word embeddings (ELMo, BERT)

• ELMo: bidirectional LSTM contextual embeddings (2018).
• BERT: transformer-based bidirectional encoder (2018-2019) that later sentence embedders build on.

Verdict in 2026: BERT-style encoders are the architectural ancestors of modern sentence embedders. For new builds, use sentence-tuned variants (SBERT, MPNet) or modern descendants.

Sentence-level embeddings (USE, SBERT, InferSent)

• Universal Sentence Encoder (USE): Google’s 2018 sentence-level embedder.
• SBERT: Sentence-BERT fine-tuned for semantic similarity.
• InferSent: Facebook’s BiLSTM-on-NLI-data sentence embedder (2017).

Verdict in 2026: Modern SBERT descendants (all-MiniLM, all-mpnet, bge-large) are the ones to try; USE and InferSent are older designs.

Universal text embeddings (E5, BGE, NV-Embed, GritLM, SFR-Embedding)

• E5: Microsoft’s general embedder family. E5-mistral-7b is the LLM-backbone variant.
• BGE: BAAI’s series; BGE-M3 is an MIT-licensed model with dense, sparse and multi-vector retrieval modes.
• NV-Embed: NVIDIA’s LLM-backbone embedder; NV-Embed-v2 ranked No. 1 on the original MTEB English leaderboard in August 2024 and ranks 24th on MTEB(eng, v2) today.
• GritLM-7B: unified generation and embedding from one model.
• SFR-Embedding-2: Salesforce’s general embedder (CC BY-NC 4.0).

Verdict in 2026: this is the category to pick from for new builds. Universal text embeddings handle retrieval, classification, clustering, and similarity from one model. Reranking can be layered on top with a cross-encoder like BGE-reranker-v2 or Cohere Rerank 4 rather than served by the same encoder.

Which Embedding Model Is Best on the MTEB Leaderboard Today?

MTEB (Massive Text Embedding Benchmark) is a public benchmark and leaderboard for embedding models, hosted on Hugging Face. Its current English board, MTEB(eng, v2), scores 41 tasks across retrieval, reranking, classification, clustering, pair classification, STS and summarization; the original English board used 56 tasks. Selected models from the MTEB(eng, v2) top 10 on October 14, 2026:

The rank uses the leaderboard’s own aggregation, so it does not follow the mean score, and models differ in how many tasks they saw in training: the zero-shot share is 48% for QZhou-Embedding, 82% for ingot-8b-r3 and 95% for Qwen3-Embedding-8B. Some models, including BGE-M3, OpenAI text-embedding-3 and Voyage 3.5, have incomplete results on this board and no comparable score. Check the live leaderboard before you pick.

What Happened to NV-Embed-v2?

NV-Embed-v2 scored 72.31 across 56 tasks and ranked No. 1 on the original MTEB English leaderboard as of August 30, 2024. On MTEB(eng, v2) it ranks 24th today with a 69.81 mean task score, and its CC BY-NC 4.0 license rules out commercial use. Other older picks sit lower too: stella_en_1.5B_v5 (69.43, rank 32), e5-mistral-7b-instruct (67.97, rank 38), SFR-Embedding-2_R (69.82, rank 42) and GritLM-7B (67.07, rank 48).

MTEB also runs language boards, so a team serving one language family can check a narrower ranking. For Scandinavian languages, MTEB(Scandinavian, v1) is led by F2LLM-v2-14B (71.10, Apache 2.0) as of October 14, 2026.

What Is the Best Multilingual Embedding Model?

On MTEB(Multilingual, v2), checked October 14, 2026, Microsoft’s harrier-oss-v1-27b leads with 74.27 (MIT license, 27B parameters, 32K tokens). Qwen3-Embedding-8B scores 70.58 (Apache 2.0); Qwen’s model card notes it ranked No. 1 on that board on June 5, 2025. gemini-embedding-001 is a closed option at 68.37 (rank 5).

BGE-M3 (MIT, 100 plus languages, 8,192 tokens) has incomplete results there and adds sparse and multi-vector retrieval to dense. Scores average many languages, so validate on a labelled set in each language you serve.

What Is the Best Small Embedding Model?

Small open models with complete MTEB(eng, v2) scores include Jasper-Token-Compression-600M (MIT, about 0.6B parameters, 74.75, rank 2, with 48% of its tasks zero-shot), Qwen3-Embedding-0.6B (Apache 2.0, 70.47, 32K tokens) and EmbeddingGemma-300m (Gemma license, 69.67, 768 dims, 2,048 tokens). For multilingual work, harrier-oss-v1-0.6b (MIT) scores 69.01 on MTEB(Multilingual, v2). For the smallest footprint, bge-small-en has about 33M parameters.

BGE-M3 vs EmbeddingGemma

BGE-M3 is MIT licensed, accepts 8,192 tokens and supports dense, sparse and multi-vector retrieval, but has no complete MTEB(eng, v2) score. EmbeddingGemma-300m is smaller, uses the Gemma license, accepts 2,048 tokens and has a complete English v2 score (69.67). Pick BGE-M3 for hybrid retrieval over long chunks, and EmbeddingGemma for a small dense model.

Capabilities and optimisation

These models are trained on broad data and fine-tuned with contrastive objectives so they handle retrieval, classification, clustering, and similarity all at once. Once a model is in production, storage and inference cost add up with every vector you store, and four techniques cut both.

Optimisation techniques that matter in production:

• Matryoshka representations: models trained this way let you truncate the embedding to its first 256 or 512 dimensions to cut storage and search cost; measure the recall change on your own set.
• Quantisation: int8 stores 4x and binary 32x less than float32 vectors, with some recall loss you should measure.
• Knowledge distillation: a smaller “student” model trained to mimic a larger “teacher”.
• Pruning: removing low-importance parameters; it can be combined with distillation.

Core architecture in modern embedding models

Encoder-only (BERT family)

BERT and its sentence-tuned descendants (SBERT, MPNet) use deep bidirectional transformer encoders to build rich contextual representations.

Decoder-only with embed pooling (LLM-backbone)

NV-Embed-v2, E5-mistral-7b, GritLM, and SFR-Embedding all start from a decoder LLM and add a pooling layer plus contrastive fine-tuning. Qwen3-Embedding follows the same pattern, built on the dense foundation models of the Qwen3 series.

Self-attention

Self-attention scores the relevance of every token to every other token. This is the mechanism that lets modern embedders capture long-range context that ELMo and BiLSTMs missed.

Pretraining objectives in 2026

• Masked language modelling (MLM) (BERT-era): predict masked tokens.
• Contrastive learning: pull similar pairs together, push dissimilar pairs apart.
• Cross-encoder pretraining: train a separate cross-encoder for reranking after dense retrieval. BGE-reranker-v2 and Cohere Rerank 4 are examples.
• Instruction tuning for embedders: prepend “Represent this for retrieval:” to the query to bias the embedding toward the task. Used by E5, NV-Embed, BGE.

Qwen3-Embedding vs BGE-M3 vs OpenAI text-embedding-3 Trade-offs

BGE-M3 is a permissively licensed starting point: it is MIT-licensed, supports 100 plus languages, and ships dense plus sparse plus multi-vector retrieval modes in one model. Pair it with a cross-encoder such as BGE-reranker-v2 when you want a sharper top-k.

How large language models power high-quality embeddings

LLM backbones now sit under many high-scoring open-weight embedders, such as Qwen3-Embedding-8B, KaLM-Embedding-Gemma3-12B and llama-embed-nemotron-8b (ranks 4, 2 and 3 on MTEB(Multilingual, v2), October 14, 2026). The training recipe:

1. Start from a pretrained decoder LLM.
2. Replace the next-token head with a pooling layer (last-token or mean-pool).
3. Contrastive fine-tune on paraphrase, NLI, and retrieval pairs.
4. Optionally add instruction prompts to bias toward specific tasks.

The result: embeddings that benefit from the LLM’s broad world knowledge, at the cost of size: NV-Embed-v2 has about 7.85B parameters, BGE-large-en-v1.5 about 335M.

How to Choose an Embedding Model in 2026

Start with two or three candidates that pass your filters, build a labelled retrieval set, and re-run the comparison as production data accumulates. Keep the same queries and chunks for every run, so a score change traces back to the model.

How to evaluate embedding models for your workload

Leaderboards lie, and your domain is different. Fix your chunking strategy first, so every candidate embeds the same chunks. Build a labelled retrieval set and score candidates yourself. For a deeper methodology on this, see evaluating embedding models in 2026.

A 5-step eval pattern:

1. Collect 200 plus queries. Pull from real production logs if available; synthetic if not.
2. Mark the correct passages. For each query, identify the chunks that should be retrieved.
3. Score recall@k. Use k=5 and k=10.
4. Score NDCG@10. Rewards correct passages near the top.
5. Add a reranker. Put BGE-reranker-v2 or Cohere Rerank 4 after retrieval, then re-score recall@k and NDCG@10.

For retrieval, Future AGI scores each query and its retrieved chunks with the built-in Context Relevance eval, which returns a score and a plain-language reason. Run every candidate embedding model on the same labelled queries and compare those scores side by side. See the built-in evals. The open-source SDK is Apache 2.0 at agent-learning-kit (pip install agent-learning-kit).

How to monitor embedding-model performance in production with Future AGI

Embedding choice shapes retrieval recall, which shapes RAG accuracy. If your embedding model drifts (or your data drifts under your embedding), your downstream eval scores drift.

Two pieces of the Future AGI platform pair with any embedding model choice:

• traceAI, Future AGI’s Apache-2.0 OpenTelemetry instrumentation, traces the retrieval steps your application runs so eval scores can sit next to each trace. Source at github.com/future-agi/traceAI.
• Evaluate scores retrieval and generation quality together with built-in evals such as Context Relevance, Chunk Attribution, Chunk Utilization and Groundedness, plus retrieval metrics Recall@K, Precision@K, NDCG@K, MRR and Hit Rate. The Apache 2.0 SDK is agent-learning-kit.

Future AGI is not an embedding model vendor: embeddings come from specialist labs (NVIDIA, BAAI, OpenAI, Voyage, Cohere, Jina). Future AGI is the eval and observability layer that pairs with whichever embedding you pick.

Shortlist Three Models, Then Let Your Own Queries Pick the Winner

A leaderboard score tells you which models deserve a test on your documents. Pick three candidates that pass your license, language and hosting filters, embed the same chunks with each, and score them on your own labelled queries with Recall@K and NDCG@K.

Run that comparison with Future AGI’s built-in retrieval evals.