Best Embedding Models in 2026

The retrieval-quality leader per dollar: four sizes sharing one embedding space, from an open-weight nano to a mixture-of-experts flagship.

The Voyage 4 series, launched January 15, 2026, is the strongest default in this roundup. Voyage reports that voyage-4-large leads the 29-dataset RTEB retrieval benchmark, beating OpenAI text-embedding-3-large by an average of 14.05 percent and Cohere Embed 4 by 8.2 percent on NDCG; those are vendor-run numbers, but the mid-tier voyage-4 at $0.06 per million tokens is the pick we actually recommend, because it approaches flagship quality at half the price of OpenAI's large model. The engineering is genuinely distinctive: all four models produce compatible embeddings in a shared space, so you can embed documents with voyage-4-large and queries with voyage-4-lite, and Voyage describes voyage-4-large as the first production-grade embedding model built on a mixture-of-experts architecture. Every model in the family takes 32,000-token inputs, offers 256 to 2,048 Matryoshka dimensions, and outputs float, int8, uint8, binary, or ubinary, which cuts vector storage bills before the database ever sees them. The trade-offs: the main models are API-only (only voyage-4-nano ships open weights, under Apache 2.0), and MongoDB ownership means the roadmap's center of gravity is Atlas.

One embedding space for text, images, audio, video, and PDFs, with MRL truncation from 3,072 down to 128 dimensions.

gemini-embedding-2 is Google's latest embedding model, billed in its docs as the first multimodal embedding model in the Gemini API, and it is the broadest model in this roundup: it embeds text, images, video, audio, and PDFs into a single semantic space, covers more than 100 languages, takes 8,192-token inputs, and uses Matryoshka Representation Learning so you can truncate from 3,072 dimensions to 128, with truncated outputs auto-normalized. For a corpus where screenshots, scanned PDFs, and recordings sit next to text, this is the strongest managed option, and the Gemini API's free tier means you can evaluate it without a card. The costs are the catch: at $0.20 per million text tokens it is the most expensive text embedding here, more than three times voyage-4, with batch halving that to $0.10. Non-text inputs meter separately: $0.45 per million image tokens, $6.50 for audio, $12.00 for video, and per-request caps apply (6 images, 180 seconds of audio, 120 seconds of video, 6-page PDFs). One structural warning from Google's own docs: embedding spaces are incompatible between model generations, so migrating from gemini-embedding-001 to 2 means re-embedding everything.

The long-document and enterprise pick: a 128,000-token context window, text plus image plus mixed-PDF inputs, and private deployment paths.

embed-v4.0's defining spec is its 128,000-token context window, sixteen times the input limit of OpenAI and Gemini and four times Voyage's. That changes pipeline design: contracts, filings, and reports can be embedded whole or in very large sections instead of being shredded into paragraph chunks, and the model natively handles text, images, and mixed text-image PDFs. Output dimensions are Matryoshka-style at 256, 512, 1,024, or 1,536. The enterprise story is the other reason it ranks here: Embed 4 is available through AWS Bedrock and through Cohere's Model Vault dedicated deployments (from $4.00 per hour or $2,500 per month), which is the path regulated buyers actually purchase. Two honest caveats. Cohere's own pricing page now leads with those dedicated instances, and per-token API pricing ($0.12 per million input tokens) is easier to find on partner platform listings than on cohere.com. And the model shipped April 15, 2025, so it is the oldest current-generation model here; Voyage's vendor-run benchmark puts voyage-4-large 8.2 percent ahead of it on retrieval.

The ecosystem default: cheap, integrated everywhere, good enough for most workloads, and visibly aging.

text-embedding-3-small at $0.02 per million tokens is the cheapest mainstream API embedding, and that plus universal ecosystem support keeps the pair on this list. Both models take 8,192-token inputs and support the dimensions parameter, OpenAI's Matryoshka-style truncation: per OpenAI's docs, text-embedding-3-large cut to 256 dimensions still outperforms the old ada-002 at full size. OpenAI reports MTEB scores of 62.3 for small and 64.6 for large. Every framework, vector database, and tutorial supports these models out of the box, which has real operational value and makes them the lowest-friction choice for a team shipping its first retrieval feature. The honest negative is momentum: OpenAI's documentation lists nothing newer than the text-embedding-3 pair, while every other vendor in this roundup has shipped a new generation since these models launched, and Voyage's vendor-run RTEB comparison claims a 14 percent retrieval gap against text-embedding-3-large. Choose OpenAI for cost and ubiquity, not for quality leadership.

The open-source workhorse: MIT license, dense plus sparse plus multi-vector retrieval from one model, and no per-token bill ever.

BGE-M3 is what you deploy when the embeddings must run on your own hardware. It is MIT licensed with no commercial restrictions, takes 8,192-token inputs, covers more than 100 languages, and its signature trick is producing three retrieval signals in one forward pass: 1,024-dimension dense vectors, sparse lexical weights, and ColBERT-style multi-vector representations. That means hybrid dense-plus-sparse retrieval, the highest-leverage retrieval upgrade for most RAG systems, comes out of a single model instead of two. Adoption is the other proof point: roughly 35 million monthly downloads on Hugging Face, which makes it the de facto open-source standard and guarantees integration support everywhere. The FlagEmbedding family also offers larger, stronger siblings (bge-multilingual-gemma2 on a gemma-2-9b base, bge-en-icl with in-context learning) when quality matters more than serving cost. The honest accounting: raw retrieval quality trails the 2026 API flagships, and free means you now own GPU serving, batching, autoscaling, and model upgrades. For sovereignty, air-gapped deployments, or embedding costs that must not scale with tokens, it is the clear pick.

Small-model engineering for edge and budget deployments: a 677M-parameter model that matches its own 3.8B predecessor on retrieval.

jina-embeddings-v5-text, released in February 2026, is the most interesting efficiency story in this roundup. The 677M-parameter small variant handles 32,000-token inputs, outputs 1,024 dimensions with Matryoshka truncation all the way down to 32, and Jina reports 67.0 on MMTEB and 71.7 on MTEB English, matching its own 3.8B-parameter v4 on retrieval at 5.6 times smaller. A 239M nano variant (8,192 tokens, 768 dimensions) goes smaller still, and the small model ships 14 GGUF quantization variants plus MLX builds, which makes this the most credible option here for on-device and edge retrieval. Language coverage is broad but reported inconsistently: Jina's model page lists 93 supported languages for the small variant while the launch post cites 119. The v5-omni siblings extend the same embedding space to text, images, audio, and video, and stay index-compatible with v5-text. The reason it ranks sixth is licensing, not engineering: the open weights are CC BY-NC 4.0, research and non-commercial only, so production use means the hosted pay-per-token API or a commercial license from Jina. That friction, plus an ecosystem far smaller than OpenAI's or BGE's, is the trade.