best 12 embedding models

Qwen3-Embedding-8B is apache-2.0 on both code and weights, which puts it in the small group here you can ship commercially without reading anything twice. it claims over a hundred languages, takes 32,768 tokens of context — four times what OpenAI or Gemini accept — and emits 4,096 dimensions that matryoshka truncation can take all the way down to 32, so the storage cost is yours to choose rather than the model's to impose.

on quality it publishes the strongest open numbers here, and unusually it labels them: 75.22 on MTEB english v2 and 70.58 on the multilingual board, per its own card. we still would not buy on that alone, but a vendor that names the board version is doing better than most of this field.

hosted, it runs about $0.010 per million tokens on DeepInfra. against OpenAI at $0.13 and Gemini at $0.20 that is not a discount, it is a different price bracket, and for a large corpus being re-embedded periodically it is the difference between a line item and a rounding error.

the cost is size. eight billion parameters and 4,096 native dimensions make it the heaviest model here to serve and, untruncated, the most expensive to index — EmbeddingGemma does the same nominal job at 300m parameters and 768 dimensions. if you are self-hosting on modest hardware, or embedding at a volume where index size dominates, the smaller open models are the more sensible pick and this one is overkill.

there is a matching Qwen3-Reranker family at 0.6b, 4b and 8b under the same apache-2.0 licence, which makes the two-stage retrieve-then-rerank pipeline buildable entirely from one permissively licensed family.

gemini-embedding-2 is Google's first multimodal embedding model, mapping text, images, video, audio and documents into a single space at over a hundred languages. that is a genuinely different capability from everything else here except Cohere, which does text and images but not audio or video. if your corpus is a mix of media and you want one index over all of it, this is the shortlist.

dimensions run from 128 to 3,072 with matryoshka support and recommended stops at 768, 1,536 and 3,072, so you can trade index size against quality deliberately. context is 8,192 tokens, four times what the older gemini-embedding-001 accepted.

the text price is $0.20 a million, the highest here, halving to $0.10 in batch mode. the other modalities are where it gets expensive in a way the headline hides: images at $0.45, audio at $6.50 and video at $12.00 per million tokens, so audio is thirty-two times the text rate and video sixty times. multimodal retrieval is a real capability and it is priced like one.

the migration is the thing to plan for. Google states that gemini-embedding-001 and gemini-embedding-2 do not produce compatible embedding spaces, so moving to the current model means re-embedding your entire corpus, not swapping a model id. anyone who built on 001 is facing a full re-index, and that cost belongs in the evaluation. we could not confirm an MTEB score for the new model — the widely quoted 68.32 belongs to 001 and should not be carried across.

this is the model most rag tutorials use and most production stacks inherited. it is good, it is well documented, and if you are already sending requests to OpenAI it costs you no new vendor relationship, no new key and no new billing conversation. $0.13 a million puts it mid-table.

its best feature is the dimensions parameter, which lets you request any smaller output size rather than picking from a fixed list the way Cohere and voyage do. shortening to 256 dimensions costs far less quality than naive truncation would, and it means index size is a dial rather than a decision made once at model selection.

the age shows in the specs. 8,192 tokens of context is the joint smallest here alongside Gemini, against 32,768 from Qwen3, voyage and Jina and 128,000 from Cohere. OpenAI publishes no language count, makes no multimodal claim, and sells no reranker, so a two-stage pipeline means bringing in a second vendor. the MTEB figure it quotes, 64.6, is unlabelled and matches its january 2024 launch benchmark, which predates the v2 english board entirely.

the straightforward read is that this is the safe institutional choice rather than the best one. nothing about it is bad. the open models beat it on price by an order of magnitude, Cohere beats it on context by a factor of fifteen, and Gemini beats it on modality — and OpenAI has not shipped a replacement in over two years.

voyage-4-large is the current flagship, having replaced voyage-3-large during the period we were researching this category — and it cut the price on the way, from $0.18 per million to $0.12. that undercuts OpenAI while offering 32,000 tokens of context against OpenAI's 8,192.

dimensions default to 1,024 with 256, 512 and 2,048 also selectable, and Voyage states that all embeddings created with the 4 series are compatible across those options, which is a more useful guarantee than it sounds — it means changing your mind about index size later does not force a full re-embed the way moving between Gemini generations does.

the reranker pricing is the quiet advantage. rerank-2.5 at $0.05 per million and a lite variant at $0.02 make the second retrieval stage cheap enough to use by default, and having both halves from one vendor removes a class of integration work. OpenAI sells no reranker at all.

what you give up is transparency. Voyage publishes no language count, no parameter count, and no MTEB score we could confirm from its own pages — a 66.8 figure circulates on aggregator sites with no board version attached, which fails our bar. and although Voyage sells separate multimodal embedding models, we could not confirm that voyage-4-large itself handles images, so treat it as text-only unless you check.

embed-v4.0 accepts 128,000 tokens in a single embedding request. that is fifteen times OpenAI's window and four times Qwen3's, and it changes what you have to do to a document before indexing it — many reports, contracts and papers fit whole, so the chunking strategy that dominates most rag design work partly stops being necessary.

it also embeds text, images and mixed text-image inputs such as pdfs into one space, which for document-heavy corpora is the practical version of multimodality — not video search, but tables and diagrams that survive as something other than lost content. dimensions are selectable at 256, 512, 1,024 or 1,536, defaulting to 1,536.

Cohere sells matching rerankers in Rerank 3.5 and Rerank 4, so the two-stage pipeline is available from one vendor here too, though those carry the same pricing problem.

and that is the problem. we fetched cohere.com/pricing twice and read the embed documentation, and there is no standard per-token rate for Embed anywhere on Cohere's public pages — only enterprise dedicated-instance pricing at $4 to $5 an hour, or $2,500 to $3,250 a month. every other vendor in this ranking, including the ones we rank below it, prints a number. we will not publish the third-party figures that circulate because we cannot source them to Cohere, and a vendor whose api cost you cannot determine before signing up is a harder sell than its specs deserve. we also found no MTEB score on Cohere's own pages, and the numbers quoted elsewhere disagree with each other by more than a point.

EmbeddingGemma is 300 million parameters, emits 768 dimensions truncatable to 512, 256 or 128, and claims over a hundred languages. it is built to run on phones and laptops, and it does. hosted on DeepInfra it costs $0.002 per million tokens — five times cheaper than the next cheapest option here and a hundredth of gemini-embedding-2 from the same company.

it publishes its MTEB numbers with the board version attached, which too few models here bother to do: 61.15 on multilingual v2 and 69.67 on english v2. those sit below Qwen3-Embedding-8B, as you would expect from a model a twenty-sixth of the size, and they are respectable for what it is.

the licence is the reason it is not higher. this is not apache-2.0 or mit — it is Google's custom Gemma Terms of Use, which is commercially usable but comes with a prohibited-use policy you must comply with, and a redistribution condition requiring you to pass the same restrictions and a copy of the agreement to anyone you give the model to. that is a materially different obligation from Qwen3 or BGE-M3, and it is worth a lawyer's five minutes if you plan to embed it in a product you ship.

the technical ceiling is context: 2,048 tokens, the shortest of any model in this ranking except the 512-token pair below it. for chunked retrieval that is workable. for anything document-length it means more chunks, more vectors and more of the boundary problems chunking creates.

BGE-M3 does something none of the others do: one model produces dense embeddings, sparse lexical weights and ColBERT-style multi-vector output. hybrid search normally means running a dense model alongside bm25 and reconciling two systems; this collapses that into one call, and for retrieval quality on mixed keyword-and-semantic queries it is a real architectural simplification.

the licence is mit, which is as permissive as this category gets, and it claims more than a hundred working languages across an 8,192-token context. it remains one of the most widely deployed open embedding models in production, and there is a matching bge-reranker-v2 family under the same licence.

the limitation is dimensional rigidity. the dense output is fixed at 1,024 with no matryoshka support, so unlike Qwen3, EmbeddingGemma or Stella you cannot trade dimensions for index size after the fact. 1,024 is a reasonable default and it is still a decision the model makes for you.

and it is getting old in a way worth naming. the model dates to january 2024, the card has not seen a substantive update since a july 2024 evaluation correction, and NVIDIA is deprecating its hosted NIM endpoint on 24 august 2026 — a hosting change rather than a model change, but the kind of signal that tends to precede others. we could not confirm a trustworthy MTEB score with an identifiable board version from its card at all.

Stella is 1.5 billion parameters on a Qwen2 backbone, mit licensed, with matryoshka dimensions at 256, 512, 768, 1,024, 2,048, 4,096, 6,144 and 8,192. that is the widest range on offer here, and the authors note that 1,024 retains nearly all the quality of 8,192 — which is a more useful statement than most vendors make, because it tells you where the knee is instead of leaving you to find it.

mit means no conditions at all: no prohibited-use policy like EmbeddingGemma, no non-commercial clause like Jina or NV-Embed, no copyleft. for a team that needs to embed a model inside a shipped product, that matters more than a point of benchmark score.

the constraints are real and they are why it sits here rather than higher. it is english only, which rules it out for most of the multilingual work the models above it handle, and the recommended input length is 512 tokens, the shortest in the ranking alongside Nomic v2. short chunks mean more vectors and more boundary loss.

there is also no hosted option we could price from a primary source, so unlike Qwen3 or BGE-M3 there is no rent-it-first path — you are provisioning gpus before you know whether it suits you. the card publishes extensive per-task MTEB results but no aggregate score with a board version we could identify, and the repository's top-level licence file 404s, so the mit claim comes from the card metadata rather than a licence file.

Nomic Embed Text v2 is a mixture-of-experts model — 475 million total parameters of which only 305 million activate on any given inference. that is an unusual design for an embedding model and it pays off exactly where you would hope: multilingual coverage of around a hundred languages at a compute cost closer to a small dense model than its total size suggests.

it is apache-2.0 on the weights, so there are no commercial conditions to work around, and it emits 768 dimensions truncatable to 256 for roughly a threefold storage reduction. Nomic has a good record of publishing training data and methodology alongside its models, which is rarer than it should be.

the killer is context: 512 tokens. that is a sixteenth of BGE-M3's window and a sixty-fourth of Qwen3's. every document of any length becomes many chunks, each chunk loses the surrounding context that would have disambiguated it, and the retrieval quality you lose to aggressive chunking is not recovered by a better embedding of each fragment. for short-text work — product titles, queries, tags, individual sentences — that is fine and the efficiency is attractive. for document retrieval it is the wrong tool.

we also could not confirm a hosted price from any primary source. it is not listed on DeepInfra, and Nomic's own pricing page did not surface a per-token embedding rate in what we could read, so treat this as a self-host proposition unless you verify otherwise.

if your stack is already Mistral, this is the frictionless option: same api, same key, same invoice, at $0.10 per million tokens, which undercuts OpenAI's $0.13. the sdk integration is clean and 8k of context is adequate for chunked retrieval. as an incremental decision for an existing customer it is easy to justify.

as a standalone choice it is hard to evaluate, because Mistral publishes remarkably little about it. we could not confirm the output dimension count from Mistral's own documentation — third-party integration docs for Pinecone and Milvus state 1,024, but those are not the vendor. there is no published parameter count, no language count, and no MTEB score on Mistral's pages. the api documents an optional output_dimension parameter described as available where the feature exists, without saying what the defaults or maximum are.

for a category where the buying decision turns on dimensions against storage cost and context against chunking strategy, not publishing the dimensions is a significant omission. every open model here tells you, and so does every other proprietary one except Cohere.

the model id is mistral-embed-2312 and it dates to december 2023, making it the oldest thing in this ranking. Mistral has since shipped a separate Codestral Embed for code retrieval, but the general-purpose embedder appears unchanged in two and a half years — a long time in this category, and unlike OpenAI's similar stasis it is not paired with a large ecosystem of tooling built specifically around it.

the engineering deserves credit. v5-text-small is 677 million parameters, takes 32,768 tokens of context, truncates from 1,024 dimensions down to 32, and claims support for 93 languages against training on 32. Jina reports 71.7 average on english MTEB, though without naming the board version. distilling v4-class quality into a sub-billion-parameter model is a real achievement.

the licence is the problem, and it is worth being blunt because the received wisdom is wrong. Jina embedding weights are non-commercial. we read v4's licence file directly: section 2(a) grants rights for non-commercial purposes only, and commercial use requires a separate agreement. v5-text-small, which replaced it in february 2026, is CC-BY-NC-4.0. neither is a licence you can build a product on without buying something extra, and Jina is routinely listed alongside apache and mit models as though it were equivalent.

the hosted api is the intended commercial path, but Jina sells it through a stripe credit top-up rather than publishing a per-million-token rate, so we cannot tell you what it costs. that is the same failure as Cohere's, on a model with a fraction of Cohere's context window.

there is also a versioning tangle to navigate. v4 was the multimodal model; the v5 line split, with v5-text-small and nano handling text and a separate v5-omni line taking over images, audio, video and pdfs from may 2026. so the model that replaced the multimodal one is not multimodal, and picking the right member of the family requires reading release notes rather than model names.

on the numbers NV-Embed-v2 is a serious model: 8 billion parameters, 4,096 dimensions, 32,768 tokens of context, and a 72.31 average across 56 MTEB tasks that took the top of the english leaderboard when it landed in 2024. if benchmark position were the only criterion it would rank near the top of this page.

it carries CC-BY-NC-4.0, and NVIDIA's model card does not leave it to inference: this model should not be used for any commercial purpose. commercial users are pointed at NeMo Retriever NIM microservices, which are a separate product under separate licensing. that is not an ambiguity to interpret in your favour — it is a sentence.

it is ranked last because of the gap between that and how it is described elsewhere. it appears on comparison pages, in framework documentation and in rag tutorials as an open model you can use, with the licence either unmentioned or described as permissive. anyone who took that at face value and embedded a production corpus with it has a problem that is not fixed by swapping the model later, because the vectors were produced by it.

the technical objections are secondary but they exist. there is no matryoshka support, so all 4,096 dimensions must be stored — the largest index of any model here, roughly four times a 1,024-dimension model on the same corpus. it is english only. and the MTEB figure, while impressive, carries no board version on the card, and the 56-task structure suggests the pre-reorganisation leaderboard rather than v2.