MTEB 2026: State of the Embeddings Benchmark

MTEB in 2026: The Landscape Has Changed

The Massive Text Embedding Benchmark (MTEB), the global reference for evaluating embedding models, has seen its rankings disrupted in 2025-2026. Open source Alibaba Qwen3 has taken the lead, Google made a splash with Gemini Embedding, and Cohere revolutionized the market with the first production multimodal embedding.

"The MTEB leaderboard constantly evolves with new submissions," explains Dr. Niklas Muennighoff, researcher at Hugging Face and creator of MTEB. "In 2026, we observe a convergence of scores between open source and proprietary APIs."

MTEB Benchmark Structure

Task Categories

MTEB evaluates embeddings across 8 main categories:

The framework covers over 1000 languages and 58 datasets for English alone.

Evaluation Metrics

The ranking uses Borda Count by default, aggregating performance across all tasks.

MTEB Ranking January 2026

Global Top 10

Evolution from 2024

OpenAI's lack of embedding updates (still text-embedding-3 from late 2023) caused them to lose the top spot.

Best Models by Category

Retrieval (document search)

Clustering

Multilingual (non-English)

To choose the right model, check our guide on choosing embeddings.

Focus: The Rise of Open Source

Qwen3 Takes the Lead

For the first time, an open source model dominates the MTEB leaderboard. Alibaba's Qwen3-Embedding-8B:

• Overall score: 70.6 (surpasses all APIs)
• License: Apache 2.0 (free commercial use)
• Size: 8B parameters
• Multilingual: Excellent on Chinese, good on European

DEVELOPERpython
from sentence_transformers import SentenceTransformer

# Load Qwen3-Embedding
model = SentenceTransformer('Alibaba-NLP/gte-Qwen3-8B-embedding')

embeddings = model.encode(
    ["Your text to encode"],
    normalize_embeddings=True
)

Implications for Businesses

This evolution changes the game:

Focus: Cohere Embed v4 and Multimodal

A Unique Innovation

Cohere Embed v4 is the only production model capable of vectorizing:

• Text
• Images
• Interleaved documents (PDFs, slides)

Its MTEB score (65.2) is lower than leaders on pure text, but it has no equivalent for visual documents.

DEVELOPERpython
import cohere

co = cohere.ClientV2('your-api-key')

# Image embedding (unique to Cohere)
response = co.embed(
    images=["data:image/jpeg;base64,..."],
    model="embed-v4",
    input_type="image",
    embedding_types=["float"]
)

For more details, see our article on Cohere Embed v4 Multimodal.

Implications for RAG Pipelines

Model Selection by Use Case

Trade-offs to Consider

Check our guide on RAG cost optimization.

Methodology and Reproducibility

How to Run the Benchmark

DEVELOPERpython
from mteb import MTEB, get_model

# Load a model
model = get_model("Alibaba-NLP/gte-Qwen3-8B-embedding")

# Run evaluation on Retrieval
evaluation = MTEB(task_types=["Retrieval"])
results = evaluation.run(model)

# Display results
print(results)

Interactive Leaderboard

The official leaderboard is available at:

• Hugging Face MTEB Leaderboard

Rankings are dynamic - new submissions can change the order at any time.

Trends Observed in 2026

1. Open Source Dominates

The gap between open source and APIs has reversed. Qwen3 surpasses OpenAI by +6 MTEB points.

2. Multimodal Emerges

Cohere paved the way. Google and OpenAI should follow in 2026-2027.

3. Domain Specialization

Specialized models (Voyage legal/finance/code) outperform generic models by 10-15% in their domains.

4. Prices Plummeting

Google Gemini Embedding at $0.008/1M tokens changes RAG economics.

Our Take

The 2026 MTEB landscape represents a turning point:

Key points:

• Open source (Qwen3) surpasses proprietary APIs
• Multimodal (Cohere v4) opens new use cases
• Prices are falling (Gemini 16x cheaper than OpenAI)

Recommendations:

• New projects: evaluate Qwen3 (performance) or Gemini (cost)
• Visual documents: Cohere Embed v4 is essential
• Existing OpenAI projects: consider migration if performance is critical

Platforms like Ailog integrate these benchmarks to automatically select the best models for your use case.

Check our detailed 2026 embedding comparison for more details.