verifier.org

NV-Embed-v2 alternatives

11 tools we tested head to head against NV-Embed-v2, ranked — and what each one actually does differently.

last reviewed 21 aug 2026 · from our best 12 embedding models ·list curated by Onur Ozcanxin

first — what you'd be leaving

NV-Embed-v2 ranks #12 of 12 in our embedding models testing. the leaderboard model you are not allowed to ship.

50/100

strong retrieval scores and an explicit ban on commercial use — ranked last because it is listed as a production option almost everywhere and it is not one.

why people look for an alternative
  • −CC-BY-NC-4.0 — the card explicitly forbids any commercial purpose
  • −widely and wrongly listed elsewhere as commercially usable
  • −no matryoshka support — all 4,096 dimensions must be stored
  • −english only, and the MTEB board version is unlabelled

stay with NV-Embed-v2 if strong published retrieval scores — 72.31 across 56 MTEB tasks is the thing you care about most — nothing below beats it on that.

the short version
best alternativeQwen3-Embedding-8Bmultilingual retrieval where the licence has to be clean and the bill has to be small89/100best free optionStellaenglish-only retrieval where you want mit weights and full control of index size68/100
advertisement
  1. 1

    Qwen3-Embedding-8B

    #1 in embedding models · apache-2.0, best open multilingual quality, and a tenth of a cent hosted

    89/100

    verdictthe best combination of open licence, multilingual quality and truncatable dimensions in the category — provided you can afford to serve an 8b model or are happy renting one.

    Qwen3-Embedding-8B vs NV-Embed-v2
     NV-Embed-v2Qwen3-Embedding-8B
    pricefree — non-commercial use only$0.010 / 1m tokens on DeepInfra
    free tieryesyes
    price / 1mfree — non-commercial only$0.010 hosted, free self-hosted
    dimensions4,096, fixed4,096, truncatable to 32
    context32,768 tokens32,768 tokens
    licenceCC-BY-NC-4.0 — non-commercialApache-2.0
    multimodalno — english text onlyno — text only

    switch formultilingual retrieval where the licence has to be clean and the bill has to be small

    pros
    • +apache-2.0 on code and weights — commercially safe without conditions
    • +32,768-token context, four times OpenAI's or Gemini's
    • +matryoshka truncation from 4,096 down to 32 dimensions
    • +publishes MTEB scores with the board version named
    • +matching apache-2.0 reranker family
    cons
    • −8b parameters — the heaviest model here to self-host
    • −4,096 native dimensions is a large index if you do not truncate
    • −text only, no multimodal path
    visit Qwen3-Embedding-8B →full Qwen3-Embedding-8B review →
  2. 2

    gemini-embedding-2

    #2 in embedding models · one vector space for text, images, audio and video

    86/100

    verdictthe only model here that puts text, images, audio and video in one shared space — at the highest text price in the ranking, and it just invalidated its own predecessor's vectors.

    gemini-embedding-2 vs NV-Embed-v2
     NV-Embed-v2gemini-embedding-2
    pricefree — non-commercial use only$0.20 / 1m text tokens
    free tieryesyes
    price / 1mfree — non-commercial only$0.20 text, $0.10 batch
    dimensions4,096, fixed128 to 3,072
    context32,768 tokens8,192 tokens
    licenceCC-BY-NC-4.0 — non-commercialproprietary api
    multimodalno — english text onlyyes — text, image, audio, video

    switch forretrieval across mixed media where text-only embeddings cannot answer the question

    pros
    • +text, images, audio, video and documents in one shared vector space
    • +100+ languages, with matryoshka dimensions from 128 to 3,072
    • +batch mode halves the text rate to $0.10 per million
    • +8,192-token context, four times its own predecessor
    cons
    • −$0.20 per million is the highest text rate in the ranking
    • −audio costs 32x and video 60x the text rate
    • −vectors are incompatible with gemini-embedding-001 — upgrading means re-embedding everything
    • −no MTEB score published for this version yet
    visit gemini-embedding-2 →full gemini-embedding-2 review →
  3. 3

    text-embedding-3-large

    #3 in embedding models · the default everyone reaches for, and it is showing its age

    83/100

    verdictstill a solid general-purpose embedding with the most flexible dimension control here — on an 8k context and a model that has not been updated since january 2024.

    text-embedding-3-large vs NV-Embed-v2
     NV-Embed-v2text-embedding-3-large
    pricefree — non-commercial use only$0.13 / 1m tokens
    free tieryesno
    price / 1mfree — non-commercial only$0.13
    dimensions4,096, fixed3,072, truncatable to any size
    context32,768 tokens8,192 tokens
    licenceCC-BY-NC-4.0 — non-commercialproprietary api
    multimodalno — english text onlyno — text only

    switch forteams already on the OpenAI api who want a known quantity and no new vendor

    pros
    • +arbitrary dimension truncation via the dimensions parameter, not a fixed list
    • +no new vendor if you are already on the OpenAI api
    • +extremely well documented and widely supported by every framework
    • +the cheap sibling, text-embedding-3-small, is $0.02 per million
    cons
    • −8,192-token context is joint smallest in the ranking
    • −no replacement shipped since january 2024
    • −no multilingual claim, no multimodal support, no reranker sibling
    • −quotes an MTEB score without naming the board version
    visit text-embedding-3-large →full text-embedding-3-large review →
    advertisement
  4. 4

    voyage-4-large

    #4 in embedding models · cheaper than OpenAI, four times the context, and a reranker to match

    82/100

    verdictbeats the default on price, context and reranking, from a smaller vendor that publishes less about how it performs.

    voyage-4-large vs NV-Embed-v2
     NV-Embed-v2voyage-4-large
    pricefree — non-commercial use only$0.12 / 1m tokens
    free tieryesyes
    price / 1mfree — non-commercial only$0.12
    dimensions4,096, fixed1,024 default; 256 to 2,048
    context32,768 tokens32,000 tokens
    licenceCC-BY-NC-4.0 — non-commercialproprietary api
    multimodalno — english text onlyunconfirmed for this model

    switch forretrieval-focused teams who want a matched embedding and reranking pair from one vendor

    pros
    • +$0.12 per million — cheaper than OpenAI and it cut price on the version bump
    • +32,000-token context, four times OpenAI's
    • +matched rerankers at $0.05 and $0.02 per million
    • +dimension options are cross-compatible within the 4 series
    cons
    • −no MTEB score published on its own pages
    • −no language count or parameter count disclosed
    • −multimodal support for this specific model is unconfirmed
    • −smaller vendor than the three above it
    visit voyage-4-large →full voyage-4-large review →
  5. 5

    Cohere Embed v4

    #5 in embedding models · 128k of context, text and pdfs in one space, and no published price

    78/100

    verdictby far the longest context here and genuine text-plus-pdf embedding — from the only vendor in this ranking that will not tell you what an api call costs.

    Cohere Embed v4 vs NV-Embed-v2
     NV-Embed-v2Cohere Embed v4
    pricefree — non-commercial use onlynot published per token
    free tieryesyes
    price / 1mfree — non-commercial onlynot published
    dimensions4,096, fixed256 / 512 / 1,024 / 1,536
    context32,768 tokens128,000 tokens
    licenceCC-BY-NC-4.0 — non-commercialproprietary api
    multimodalno — english text onlyyes — text, images, pdfs

    switch forlong-document and pdf retrieval where a 128k window removes the chunking problem

    pros
    • +128,000-token context — fifteen times OpenAI's, the longest here by far
    • +text, images and pdfs embedded into one space
    • +matching Rerank 3.5 and Rerank 4 models
    • +four selectable output dimensions from 256 to 1,536
    cons
    • −no per-token api price published anywhere on cohere.com
    • −only dedicated-instance pricing is public, from $2,500 a month
    • −no MTEB score on its own pages, and third-party figures disagree
    • −no language count stated for v4.0 specifically
    visit Cohere Embed v4 →full Cohere Embed v4 review →
  6. 6

    EmbeddingGemma

    #6 in embedding models · the cheapest way to embed anything, if the gemma terms suit you

    75/100

    verdict300m parameters that run on a laptop at a fiftieth of Gemini's price — held back by a licence that is commercially usable but not open source.

    EmbeddingGemma vs NV-Embed-v2
     NV-Embed-v2EmbeddingGemma
    pricefree — non-commercial use only$0.002 / 1m tokens on DeepInfra
    free tieryesyes
    price / 1mfree — non-commercial only$0.002 hosted, free self-hosted
    dimensions4,096, fixed768, truncatable to 128
    context32,768 tokens2,048 tokens
    licenceCC-BY-NC-4.0 — non-commercialGemma Terms of Use
    multimodalno — english text onlyno — text only

    switch foron-device and high-volume embedding where size and cost dominate everything else

    pros
    • +$0.002 per million hosted — the cheapest rate in the ranking
    • +300m parameters, genuinely runs on a laptop or phone
    • +matryoshka truncation to 512, 256 or 128 dimensions
    • +publishes MTEB scores with the board version named
    cons
    • −Gemma Terms of Use, not an osi licence — prohibited-use policy attached
    • −redistribution obliges you to pass the same restrictions downstream
    • −2,048-token context is among the shortest here
    • −quality sits below the larger open models, as expected at 300m
    visit EmbeddingGemma →full EmbeddingGemma review →
  7. 7

    BGE-M3

    #7 in embedding models · mit, three retrieval modes in one model, and quietly ageing

    73/100

    verdictthe most versatile open model here — dense, sparse and multi-vector output under mit — on a card that has barely moved since 2024.

    BGE-M3 vs NV-Embed-v2
     NV-Embed-v2BGE-M3
    pricefree — non-commercial use only$0.010 / 1m tokens on DeepInfra
    free tieryesyes
    price / 1mfree — non-commercial only$0.010 hosted, free self-hosted
    dimensions4,096, fixed1,024, fixed
    context32,768 tokens8,192 tokens
    licenceCC-BY-NC-4.0 — non-commercialMIT
    multimodalno — english text onlyno — text only

    switch forhybrid retrieval where you want dense and sparse vectors from a single model

    pros
    • +mit licensed — the most permissive terms in the ranking
    • +dense, sparse and multi-vector retrieval from one model
    • +100+ languages at 8,192 tokens of context
    • +matching bge-reranker-v2 family under the same licence
    cons
    • −fixed 1,024 dimensions with no matryoshka truncation
    • −card essentially unchanged since july 2024
    • −NVIDIA NIM is deprecating its hosted endpoint in august 2026
    • −no MTEB score with an identifiable board version
    visit BGE-M3 →full BGE-M3 review →
  8. 8

    Stella

    #8 in embedding models · mit weights with dimensions from 256 to 8,192, english only

    68/100

    verdictthe widest dimension range in the category under an unrestricted mit licence — english-only, short-context, and with nobody hosting it for you.

    Stella vs NV-Embed-v2
     NV-Embed-v2Stella
    pricefree — non-commercial use onlyfree — self-host only
    free tieryesyes
    price / 1mfree — non-commercial onlyfree — self-host only
    dimensions4,096, fixed256 to 8,192
    context32,768 tokens512 tokens recommended
    licenceCC-BY-NC-4.0 — non-commercialMIT
    multimodalno — english text onlyno — text only

    switch forenglish-only retrieval where you want mit weights and full control of index size

    pros
    • +mit licensed with no usage conditions whatsoever
    • +eight matryoshka dimension options from 256 to 8,192
    • +the authors state where quality plateaus, rather than leaving you to test
    • +1.5b parameters — far lighter to serve than Qwen3-8B
    cons
    • −english only
    • −512-token recommended input, the shortest here
    • −no hosted option we could price — self-host or nothing
    • −licence comes from card metadata; the repo licence file 404s
    visit Stella →full Stella review →
  9. 9

    Nomic Embed Text v2

    #9 in embedding models · a mixture-of-experts embedder that only wakes up two thirds of itself

    66/100

    verdicta clever, genuinely open, efficient multilingual model undone for most rag work by a 512-token context.

    Nomic Embed Text v2 vs NV-Embed-v2
     NV-Embed-v2Nomic Embed Text v2
    pricefree — non-commercial use onlyfree — self-host
    free tieryesyes
    price / 1mfree — non-commercial onlyfree — self-host
    dimensions4,096, fixed768, truncatable to 256
    context32,768 tokens512 tokens
    licenceCC-BY-NC-4.0 — non-commercialApache-2.0
    multimodalno — english text onlyno — text only

    switch forself-hosted multilingual embedding on constrained compute

    pros
    • +apache-2.0 weights with no commercial conditions
    • +mixture-of-experts design activates only 305m of 475m parameters
    • +around 100 languages from a very small model
    • +matryoshka truncation from 768 to 256 dimensions
    cons
    • −512-token context makes it unsuitable for document retrieval
    • −no hosted list price confirmable from a primary source
    • −no MTEB score published on the model card
    • −no reranker sibling
    visit Nomic Embed Text v2 →full Nomic Embed Text v2 review →
  10. 10

    Mistral Embed

    #10 in embedding models · a flat price, a tidy api, and almost no published specification

    62/100

    verdictcheaper than OpenAI and easy to adopt if you are already a Mistral customer — from a model that publishes neither its dimensions nor its benchmark scores.

    Mistral Embed vs NV-Embed-v2
     NV-Embed-v2Mistral Embed
    pricefree — non-commercial use only$0.10 / 1m tokens
    free tieryesyes
    price / 1mfree — non-commercial only$0.10
    dimensions4,096, fixednot published by vendor
    context32,768 tokens8,000 tokens
    licenceCC-BY-NC-4.0 — non-commercialproprietary api
    multimodalno — english text onlyno — text and code

    switch forteams already building on Mistral who want embeddings on the same bill

    pros
    • +$0.10 per million, cheaper than OpenAI's flagship
    • +no new vendor if you are already using the Mistral api
    • +flat pricing with no tiers or dimension-based surcharges
    • +a separate Codestral Embed exists for code retrieval
    cons
    • −output dimensions are not published by the vendor
    • −no parameter count, language count or MTEB score published
    • −dates to december 2023 — the oldest model in this ranking
    • −no weights, no self-host path, no reranker sibling
    visit Mistral Embed →full Mistral Embed review →
+ 1 more tested, not detailed here
we ranked 12 embedding models in total. the 1 that didn't make this page are written up in the full ranking →

how these were compared

every tool on this page went through the same test as NV-Embed-v2 — same tasks, same order, scored the same way. the comparison tables are the figures from that testing, not vendor spec sheets.

the embedding models test in full →
was this useful?