Best open-source embedding models in 2026
Embedding models turn text, and now images, video and audio, into vectors for search and retrieval-augmented generation. Most are small enough to run on a laptop CPU, so license, language coverage and input length matter more than raw size. Look here before you pay for an embedding API.
-
BGE-M3 Most downloaded
BAAI's MIT-licensed model with dense, sparse and multi-vector search in one model, 100+ languages and 8,192-token inputs. It had about 29 million downloads in the 30 days to 11 November 2026, far more than any other embedding model here.
-
Qwen3-Embedding
Alibaba's Apache 2.0 family in 0.6B, 4B and 8B sizes: 100+ languages, 32K context and adjustable vector length. The 0.6B model had about 8.9 million downloads in the 30 days to 11 November 2026.
-
Harrier OSS v1
Microsoft's MIT-licensed family in 270M, 0.6B and 27B sizes, with 32K-token inputs and 42 named languages. Microsoft reports Multilingual MTEB v2 scores of 66.5, 69.0 and 74.3, which are self-reported, and the 0.6B had about 222,000 downloads in the 30 days to 11 November 2026. Queries need a one-sentence task instruction.
-
EmbeddingGemma 2
Google DeepMind's 740M Apache 2.0 model puts text, code, images, video and audio in one search space and runs on-device.
-
Nemotron 3.5 Embed Newest
NVIDIA's 8B and 1B retrieval models search document page images and text without running OCR first, with a matching 8B reranker. NVIDIA reports 61.17 nDCG@10 for the 8B on ViDoRe v3.1, a self-reported score. OpenMDW-1.1 allows commercial use.
-
Qwen3-VL-Embedding
Alibaba's Apache 2.0 multimodal pair (2B and 8B) puts text, images, screenshots and video in one search space, so you can search documents that mix words and pictures.
-
Granite Embedding R2 Best small model
IBM's Apache 2.0 ModernBERT models from 47M to 311M parameters, for English or 52 enhanced languages plus code. They are small enough for a CPU, and IBM says the training data uses only permissive, enterprise-friendly licenses.
-
voyage-4-nano
Voyage AI's 346M Apache 2.0 embedder, whose vectors share a space with the hosted Voyage 4 models, so you can build locally and switch later without re-indexing. The blog gives no score for the nano model, and the BF16 file is 0.69 GB.
-
Jina Embeddings v5
Jina AI's v5 pair of 677M and 239M multilingual models, distilled from Qwen3-Embedding-4B. They are free for non-commercial use only under CC BY-NC.
-
Jina Embeddings v5 Omni
Jina's small and nano models that put text, images, video and audio in one space, keeping text vectors identical to v5 text so an existing index needs no rebuild. CC BY-NC means a commercial product needs a separate license, and the scores are Jina's own.
-
pplx-embed
Perplexity's MIT-licensed family: dense 0.6B and 4B text models with 32K context and native int8 vectors, contextual models, and v2 multimodal late-interaction and contextual models up to 9B. The v1 0.6B had about 64,700 downloads in the 30 days to 11 November 2026, and the v2 scores are Perplexity's own.
-
Giga-Embeddings
Sber's MIT-licensed Russian and English embedders in 480M, 3B and 10B mixture of experts sizes. The 480M model scores 70.98 on MTEB Russian in Sber's table, ahead of Qwen3-Embedding-0.6B, which leads it on English, code and multilingual tests.
How we chose
- Open weights you can download today from the developer's official repository.
- Languages, maximum input length and modalities are listed, because they decide whether a model fits your documents.
- Scores on benchmarks such as MTEB are self-reported and easy to tune for, so test a few models on a few hundred real queries from your own data.
- We read each license, because the Jina models here are free for non-commercial use only.