text-embedding-3-small

Overview. text-embedding-3-small is OpenAI's third-generation text embedding model, released in January 2024 alongside the larger text-embedding-3-large. It converts text into 1,536-dimension vectors by default (with configurable dimension truncation for smaller storage footprints) and is available exclusively through OpenAI's API — there is no open-weight release. Developers reach for it when they need a compact, general-purpose text encoder for semantic search, clustering, or retrieval-augmented generation pipelines.

Key strengths. As a pure text-to-vector encoder, it performs a single forward pass per input rather than an autoregressive decode loop, which supports low end-to-end latency in real-time retrieval pipelines. It supports input sequences up to roughly 8,191 tokens using the cl100k_base tokenizer, and its configurable output dimensionality lets teams trade vector size against storage and search-index cost without re-embedding from scratch. No independent quality benchmarks (retrieval accuracy, MTEB-style scores) are available in current research, so strengths here are architectural rather than benchmark-verified.

When to use it. Use this model for semantic search, document clustering, deduplication, or as the retrieval layer in a RAG pipeline where you need fast, single-pass text encoding rather than generative output. It's a reasonable default when you want a smaller, faster-to-produce embedding than OpenAI's larger third-generation sibling and don't need the absolute highest retrieval fidelity available in that family. It is not a generative or chat model — it has no system-prompt, tool-calling, or streaming support, so it should sit purely in an embedding/indexing stage of an application.

Limitations. OpenAI has not disclosed the model's architecture, parameter count, or training-data cutoff, and there is no public weights repository — it is closed and API-only. The 8,191-token input cap is modest compared to modern generative context windows, and output is a fixed-dimension vector rather than text, so it cannot be used for any generative task.

Sovereignty note. OpenAI is a US-based creator with confirmed CLOUD Act exposure, and text-embedding-3-small has no self-hosted Sovereign alternative — though eligible API customers can opt into EU data residency, teams routing regulated or sensitive data should still treat this as a non-Sovereign, US-jurisdiction endpoint given OpenAI's CLOUD Act exposure.