OpenAI: Text Embedding 3 Large (batch)

OpenAI: Text Embedding 3 Large (batch) is an embedding model from OpenAI. It sits in the embeddings category, and OpenAI describes text-embedding-3-large as its most capable embedding model for both English and non-English tasks. The batch variant exposed on api14 uses the model id openai/text-embedding-3-large:batch and was released by the provider on 2025-10-30.

The model takes text as input and returns embeddings as output. An embedding is a numerical representation of text that can be used to measure the relatedness between two pieces of text. The context window is 8192 tokens, and the maximum output is 7372, which reflects the size of the vector the model produces rather than generated prose. There is no text generation here: you send text in and get vectors back.

Because it returns vectors rather than answers, the practical uses follow from comparing those vectors. Semantic search is the obvious one: embed your documents and your query, then rank documents by similarity. The same mechanism supports retrieval for RAG pipelines, where embeddings select the passages that a separate generation model will read. Clustering and topic discovery work by grouping texts whose embeddings sit close together. Deduplication and near-duplicate detection use the same distance comparisons, and classification can be done by embedding labelled examples and assigning new texts to the nearest group. The stated strength across English and non-English tasks makes it a reasonable choice for multilingual corpora, and the 8192 token window allows fairly long passages to be embedded in one call instead of being split.

You call it through api14's OpenAI-compatible API using the model id openai/text-embedding-3-large:batch. Because the surface is OpenAI-compatible, you can point an existing OpenAI client at the api14 base URL and send your input as you normally would to an embeddings endpoint. Note that this model lists no supported parameters beyond the standard request fields, so do not expect tuning knobs such as custom dimensions or provider-specific options; check the live price table on the page for current rates rather than assuming any pricing here.

One limit worth planning around is the input side. Text is the only accepted modality, so images, audio and other formats need a different model or a preprocessing step. The model also produces embeddings only, so it cannot summarise, answer questions or follow instructions; pair it with a generation model when you need those. If your task is generation, reasoning or tool use rather than similarity and retrieval, another model in the catalog will fit better.