Text Embeddings Inference

license

Apache-2.0(OSI)

source

public(the Rust router and model backends build from the repository)

core features withheld

no — no enterprise path in the repository tree

history

HFOIL 1.0 through v1.2.0(Apache-2.0 from v1.2.3)

Text Embeddings Inference is Apache-2.0 and builds whole from the public repository, with no enterprise edition or license key, and Hugging Face's paid Inference Endpoints run it as one engine among several rather than holding features back. Its early releases used Hugging Face's restrictive HFOIL license until the v1.2 series.

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference@v1.2.0/LICENSE recorded 2026-10-13

  The LICENSE at the v1.2.0 tag is a different document, headed "Hugging Face Optimized Inference License 1.0 (HFOILv1.0)".

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference@v1.2.3/LICENSE recorded 2026-10-13

  The LICENSE at the v1.2.3 tag is the Apache License, Version 2.0, the same text the current release carries.

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference/ recorded 2026-10-13

  The repository tree at v1.9.4 holds backends, core, router, proto, docs and test folders with Cargo.toml, per-platform Dockerfiles (CUDA, ROCm, Intel, arm64) and one LICENSE; there is no ee/, enterprise/, commercial/ or proprietary/ path.

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference/LICENSE recorded 2026-10-13

  LICENSE at the latest release (v1.9.4) is the Apache License, Version 2.0, January 2004, ending "Copyright 2022 Hugging Face".

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference/README.md recorded 2026-10-13

  The README documents running the server from Docker images at ghcr.io/huggingface/text-embeddings-inference, a local install, "brew install text-embeddings-inference" on Apple Silicon and Docker builds from source, with no paid tier or license key.

• https://huggingface.co/docs/inference-endpoints/engines/tei recorded 2026-10-13

  Hugging Face's Inference Endpoints documentation lists Text Embeddings Inference (TEI) as one of its inference engines beside vLLM, TGI, SGLang and llama.cpp, configured with the same batch and concurrency settings the open server exposes.

TEI ships mainly as Docker images on GitHub's container registry, which publishes no pull counts, so GitHub stars are the only signal that covers it. Its Homebrew build for Apple Silicon is countable but is a minor install route and sees only a few dozen installs a month, so it is not used to judge reach.

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference/README.md recorded 2026-10-13

  The README's Docker Images table lists every supported backend as an image under ghcr.io/huggingface/text-embeddings-inference, from cpu-1.9 to hopper-1.9; Homebrew is offered only for Apple Silicon.

• https://formulae.brew.sh/api/formula/text-embeddings-inference.json recorded 2026-10-13

  Homebrew analytics for the text-embeddings-inference package: 49 installs in the last 30 days and 445 in the last 365 days.

• https://img.shields.io/github/stars/huggingface/text-embeddings-inference.json recorded 2026-10-13

  shields.io relay of the GitHub stargazer count for huggingface/text-embeddings-inference: 5.1k stars.

Text Embeddings Inference covers what search and retrieval need from a server, with dynamic batching, optimized kernels and production tracing across CPU, NVIDIA, AMD and Apple hardware, and it is still actively released. It sits level with its sibling Text Generation Inference as a focused server rather than a general-purpose engine.

• https://cdn.jsdelivr.net/gh/huggingface/text-embeddings-inference/README.md recorded 2026-10-13

  The README lists "Token based dynamic batching", Flash Attention, Candle and cuBLASLt kernels, Safetensors and ONNX weight loading, "Metal support for local execution on Macs" and OpenTelemetry tracing with Prometheus metrics; it documents re-rankers, sequence classification, SPLADE pooling, gRPC, AMD Instinct GPUs via ROCm, and images from Turing to Blackwell.

• https://huggingface.co/docs/inference-endpoints/engines/tei recorded 2026-10-13

  The Inference Endpoints TEI page describes it as a production-ready engine for text embeddings with token-based dynamic batching, Flash Attention, Candle and cuBLASLt, Safetensors and ONNX formats, and settings for max tokens per batch and max concurrent requests.