text-embedding-3-large
Text Embedding 3 Large: OpenAI's Best Embedding Model
Model Overview
text-embedding-3-large is OpenAI's most capable text embedding model, released in January 2024. It produces 3072-dimensional embeddings with significantly improved performance over the previous Ada-002 embedding model — achieving 54.9% on MTEB (Multilingual Text Embedding Benchmark) versus Ada-002's 31.4%. It supports shortening embeddings to reduce storage/computation costs without significant performance loss.
✨ Key Features
| Feature | Description |
|---|---|
| 3072 Dimensions | High-dimensional embeddings for maximum semantic richness |
| Flexible Dimensions | Supports reducing to any dimension ≤3072 via dimensions parameter |
| MTEB Performance | 54.9% on MTEB — 75% improvement over text-embedding-ada-002 |
| Multilingual | Strong multilingual embedding performance across many languages |
| Cost-Effective | Significantly cheaper than Ada-002 per token |
📊 Benchmarks
| Benchmark | text-embedding-3-large | text-embedding-ada-002 |
|---|---|---|
| MTEB Overall | 54.9% | 31.4% |
| Avg Retrieval (BEIR) | 54.9% | 49.3% |
🔗 Resources
| Resource | Link |
|---|---|
| API Docs | platform.openai.com/docs/guides/embeddings |
| Announcement | OpenAI Blog |
📜 License & Access
Proprietary — Available via OpenAI API.
Key Features
Supports Matryoshka Representation Learning for truncated output dimensions
8,192 token context window
Highly cost-efficient API pricing
Optimized for vector database search and RAG
You might also want to compare
Verified Sources
Tags
Model Specs
Parameters
Undisclosed
Context Window
8192
License
Proprietary
Deployment
Resources & Links
Lineage
Predecessor
text-embedding-ada-002Curator Notes
Features Matryoshka learning allowing vectors to be truncated without severe performance drop.
Compare Specs
Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.
Compare Model