OpenAI text-embedding-3-large

Acerca de OpenAI

Fundado en 2015 · San Francisco, California, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman, and restructured to capped-profit OpenAI LP in 2019. The embedding model family started with text-embedding-ada-001 in 2021, was unified into text-embedding-ada-002 in December 2022 (still the most-used embedding model on Earth at one point) and replaced in January 2024 by text-embedding-3-small and text-embedding-3-large. The v3 release was OpenAI's first to support Matryoshka-style dimension reduction (sale of arbitrary 256-3072 dim vectors from the same model) and beat ada-002 by 20+ percentage points on the MIRACL multilingual retrieval benchmark while costing roughly the same.

Arquitectura

Transformer bi-encoder with Matryoshka representation learning

OpenAI text-embedding-3-large is the flagship embedding model in the v3 generation. It produces 3,072-dimensional vectors by default with full Matryoshka representation support, so callers can request any dimension between 256 and 3,072 in the API and OpenAI will truncate-and-renormalise without re-running the model. The model accepts up to 8,191 tokens per input and was trained with a contrastive retrieval objective on a curated multilingual web corpus, including search-query/document pairs and large-scale instruction-tuned pairs. It scores 64.6% on MTEB and 54.9% on MIRACL multilingual retrieval, a step up from 61.0% / 31.4% for ada-002. Pricing is $0.00013 per 1k tokens. Output vectors are L2-normalised. Like all OpenAI models the system is closed-source and hosted only. OpenAI has not published a technical paper for the v3 family beyond a launch blog.

Parámetros

Undisclosed

Contexto

8,191 tokens

Capacidades

• 3,072-dim vectors with Matryoshka truncation to any 256-3072 size
• 8,191-token context window for long-document embedding
• Multilingual coverage across 100+ languages
• State-of-the-art MIRACL multilingual retrieval (~55%)
• L2-normalised vectors with cosine similarity
• Drop-in upgrade from ada-002 via the same /embeddings endpoint
• Best for: production RAG, multilingual search, retrieval-heavy SaaS

Entrenamiento y licencia

Not disclosed. OpenAI describes a 'curated multilingual web corpus' with search-query/document pairs and instruction-tuned pairs.

Licencia: Proprietary commercial API. Generated embeddings may be stored and used commercially under the OpenAI Usage Policy.

Pruebas de seguridad: Embeddings are not subject to content-filter constraints; OpenAI publishes general bias and fairness statements for the embeddings endpoint.