OpenModelsHQ

Embedding models 24

We track 24 open embedding models from 8 labs, 8 of them released in the 12 months to 7 Nov 2026. The newest is embeddinggemma-2 (14 Sep 2026) and the most downloaded is Qwen3-Embedding-0.6B, with 8,620,434 downloads in 30 days. 19 fit a 16 GB GPU at 4-bit, and none has a hosted API in our data.

Embedding models turn inputs into vectors that can be compared for search, retrieval and clustering. Sentence Transformers supports building and serving these models; MTEB evaluates their behaviour across tasks. A good retrieval result depends on the corpus and query distribution as well as the model.

What the category covers

The core job is turning text into a fixed-length vector for similarity comparison. Around that sit several related jobs that use the same machinery.

Rerankers are one example. A retrieval system typically fetches many candidates by vector similarity, then reranks the shortlist with a model that looks at the query and document together, which adds query-time computation and can improve relevance. Repository names carrying rerank are doing that job, not the first-stage one. Classification models, reward models and token-level variants also sit in this family, as related non-generative tasks, although their architectures and outputs differ.

Some models in our embedding category are encoders for other modalities, including audio and code. The naming is the quickest signal.

How they are measured

MTEB is the shared benchmark, and its own description has widened: a multimodal toolbox for evaluating embeddings and retrieval systems, installed with pip install mteb, with a leaderboard hosted on Hugging Face.

MTEB spans task types rather than one, so a model can be strong at retrieval and weaker at clustering or classification. The leaderboard lets you filter to the tasks and languages you care about, and a model at the top of the overall table may be beaten on your language or your domain by something smaller.

A public leaderboard is a starting point for evaluation. Test representative queries against your own documents, including difficult negatives and the languages your users write. An aggregate benchmark result need not predict the ordering on that workload.

What serves them

Sentence Transformers provides tools to compute embeddings and train embedding and reranker models. Its README documents installation with pip install -U sentence-transformers, supported Python and PyTorch versions, and optional ONNX and OpenVINO backends. The model card determines the correct input formatting and whether a query needs a special prompt.

vLLM can serve these models too, in what it calls pooling mode. Its pooling documentation says pooling models are supported primarily for convenience and that this is not guaranteed to provide performance improvements over using Hugging Face Transformers or Sentence Transformers directly, with optimisation listed as planned work. Because some architectures support both generative and pooling tasks, the supported-models doc says to specify --runner pooling explicitly so the model is used in pooling mode instead of generative mode. Separate pages cover classification, embedding, reward, scoring and token-level usage.

For embeddings at scale, the vector store matters as much as the model, and that is a choice outside the model itself.

The hardware reality

Hardware still matters for embedding workloads. Model size, input length and batch size affect memory and throughput. A CPU deployment may suit occasional queries while large indexing jobs benefit from batching on an accelerator. Measure the actual documents and queries before sizing a service.

The costs that matter are elsewhere. Embedding dimension drives storage and search cost across the whole corpus, so a larger vector multiplies your index, and some model families support truncating dimensions to trade accuracy for size. Maximum sequence length determines how you chunk documents. Reranking is the expensive step at query time, because it runs the model once per candidate pair rather than once per document.

The operational constraint people hit is not speed but consistency: changing embedding model means re-embedding the entire corpus, since vectors from different models are not comparable.

Licence patterns

Embedding checkpoints can carry permissive or custom terms. Our Apache 2.0, MIT and comparison pages explain common permissive licences.

A library’s licence is separate from the licences of the checkpoints it loads. Some model families have their own conditions, such as those described on our Gemma terms page. Check the actual checkpoint’s licence and any incorporated base model before commercial use.

This summarises licence text. It is not legal advice.

FAQ

Is a higher MTEB score always better for my use case?

No. MTEB spans multiple task types and languages, and the overall ranking can hide weakness on the task you need. Filter the leaderboard, then test on your own data.

Should I serve embeddings with vLLM?

Only if it simplifies your stack. vLLM’s own documentation says pooling support exists primarily for convenience and is not guaranteed to beat Sentence Transformers or Transformers directly.

What is the difference between an embedding model and a reranker?

An embedding model encodes each text independently so vectors can be compared cheaply at scale. A reranker scores a query and a candidate together, which adds computation per candidate and can improve relevance, so it typically runs on a shortlist.

Can I switch embedding models later?

Usually this requires rebuilding the index with the new model. Independently trained encoders do not generally produce compatible vector spaces; retain the old encoder until migration is complete.

Pick a view

All models

Newest

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
14 Sep 2026Embedding744M0.6 GBFits517,987No tracked API
25 Aug 2026Embedding2.7B1.9 GBFits12,392No tracked API
14 Jul 2026Embedding8B6.5 GBFits128,586No tracked API
14 Jul 2026Embedding1.1B1.3 GBFits324,819No tracked API
20 Apr 2026Embedding312M0.8 GBFits83,891No tracked API
20 Apr 2026Embedding97M0.2 GBFits188,743No tracked API
30 Mar 2026Embedding596M1.4 GBFits217,416No tracked API
10 Mar 2026Embedding2.3B2.2 GBFits133,622No tracked API
4 Nov 2025Embedding11.8B8.4 GBFits11,560No tracked API
16 Oct 2025Embedding1.2B1.1 GBFits297,423No tracked API
16 Oct 2025Embedding1.2B1.1 GBFits350,783No tracked API
29 Sep 2025Embedding474M1.3 GBFits62No tracked API

Most downloaded

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
3 Jun 2025Embedding596M1.4 GBFits8,620,434No tracked API
17 Jul 2025Embedding48M0.2 GBFits6,828,205No tracked API
2 Mar 2022EmbeddingUnknownUnknownUnknown1,359,122No tracked API
17 Jul 2025Embedding303M0.2 GBFits1,085,038No tracked API
16 Feb 2023EmbeddingUnknownUnknownUnknown679,663No tracked API
14 Sep 2026Embedding744M0.6 GBFits517,987No tracked API
2 Mar 2022EmbeddingUnknownUnknownUnknown501,075No tracked API
16 Oct 2025Embedding1.2B1.1 GBFits350,783No tracked API
14 Jul 2026Embedding1.1B1.3 GBFits324,819No tracked API
16 Oct 2025Embedding1.2B1.1 GBFits297,423No tracked API
30 Apr 2024EmbeddingUnknownUnknownUnknown231,063No tracked API
30 Mar 2026Embedding596M1.4 GBFits217,416No tracked API

Smallest that fit 16 GB

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
17 Jul 2025Embedding48M0.2 GBFits6,828,205No tracked API
20 Apr 2026Embedding97M0.2 GBFits188,743No tracked API
2 Mar 2022Embedding125M0.4 GBFits697No tracked API
2 Mar 2022Embedding125M0.4 GBFits9,005No tracked API
4 Dec 2024Embedding125M0.4 GBFits208,133No tracked API
4 Dec 2024Embedding278M0.5 GBFits72,421No tracked API
17 Jul 2025Embedding303M0.2 GBFits1,085,038No tracked API
20 Apr 2026Embedding312M0.8 GBFits83,891No tracked API
29 Sep 2025Embedding474M1.3 GBFits62No tracked API
3 Jun 2025Embedding596M1.4 GBFits8,620,434No tracked API
30 Mar 2026Embedding596M1.4 GBFits217,416No tracked API
14 Sep 2026Embedding744M0.6 GBFits517,987No tracked API
Top 12 in each view. Memory is the 4-bit estimate with an 8k context; a model fits when it stays within 90% of the GPU's memory.How we calculate

Releases per month

8 in 12 months
Embedding model releases per month
Months run to the newest release in our data, across all model types. Small screens show the last 6 months; the data table has all 12.
Embedding model releases per month: data
MonthTotal
Dec 20250
Jan 20260
Feb 20260
Mar 20262
Apr 20262
May 20260
Jun 20260
Jul 20262
Aug 20261
Sep 20261
Oct 20260
Nov 20260

Cheapest hosted APIs

All API prices

We have not recorded API pricing for these embedding models yet.

Top labs for embedding

Most tracked embedding models first

Sources

Other model types