Hayz Labs · https://hayzlabs.com/reviews/top/embedding-models/
reviews · top 20 list · checked 18 Oct
Top 17 embedding models
Models that turn text into numbers so software can search by meaning: the engine behind document search and chat over your files.
Nothing on this list is scored yet: no outside source we can turn into a number covers them. The order below follows the public measure named on each row.
Ordered by the MTEB leaderboard, an independent test of search and retrieval quality. One row per model line, newest version only. Open the Table to see the number that placed each one and how it was scored.
- 1harrier-oss-v1-27bMicrosoft
Top open embedding model on the multilingual board; big (27B) so needs a strong GPU. 270M and 0.6B sizes exist
Best for Teams that want the top open search model and can host it themselvesYou give up Needs an estimated 54 GB of memory at full size; plan on a data-center GPU
Not scored yetFree - 2KaLM-Embedding-Gemma3-12BTencent
Strong open multilingual search model for teams that can host a 12B model
Best for Multilingual search on your own servers, a step down in size from the top pickYou give up Needs an estimated 24 GB of memory; scores below harrier-oss-v1-27b
Not scored yetFree - 3llama-embed-nemotron-8bNVIDIA
NVIDIA's open 8B embedding model; check its licence before commercial use
Best for Multilingual search on a single mid-size GPU, ranked third on the boardYou give up Licence may rule out commercial use; needs an estimated 16 GB of memory
Not scored yetFree - 4Qwen3-Embedding-8BAlibaba
Popular Apache-licensed open model; 4B (rank 6) and 0.6B (rank 20) sizes suit smaller machines
Best for Commercial search projects that want a free model under a permissive licenceYou give up Needs an estimated 16 GB at full size; the smaller sizes rank lower
Not scored yetFree - 5gemini-embedding-001Google
Best-ranked hosted API on the board; Google keeps it for text-only use next to Gemini Embedding 2
Best for Teams that want hosted text embeddings rather than running their own modelYou give up Price not in our files; Google plans to shut it down on 14 May 2028
Not scored yetNo public price - 6Octen-Embedding-8BOcten
Open Apache-licensed model built on Qwen3-Embedding-8B and tuned further for search
Best for A free, permissively licensed alternative to Qwen3-Embedding-8B for searchYou give up Ranks below the Qwen3 model it is built on; needs an estimated 16 GB of memory
Not scored yetFree - 7F2LLM-v2-14BAnt Group
Open family from 330M to 14B; the 8B (rank 9) and 4B (rank 12) are lighter options
Best for Teams that want one model family in several sizes to match their hardwareYou give up The 14B needs an estimated 28 GB of memory; the lighter sizes rank lower
Not scored yetFree - 8Seed1.6-embeddingByteDance
ByteDance's hosted embedding API; a fit mainly for teams already on ByteDance's cloud
Best for Hosted multilingual search without running your own serversYou give up Price not in our files; ranks below several free open models on the board
Not scored yetNo public price - 9Giga-Embeddings-instruct-10Bai-sage (Sber)
Open model that runs only 1.8B of its 10B parameters at a time; a 3B size ranks 18
Best for Teams hosting their own search who want less compute per requestYou give up Still needs an estimated 20 GB of memory to load all 10B parameters
Not scored yetFree - 10jina-embeddings-v5-text-smallJina AI (Elastic)
Small 0.6B model that ranks near much bigger ones; free weights are for non-commercial use
Best for Research and personal projects needing a small multilingual model on modest hardwareYou give up Free weights are non-commercial only; business use means paying for its API
Not scored yetFree - 11BidirLM-Omni-2.5B-EmbeddingBidirLM
Apache-licensed mid-size open model; a 1.7B text version ranks 24
Best for Free self-hosted search on machines with around 5 GB of spare memoryYou give up Ranks 17th on the board, behind the larger open models above it
Not scored yetFree - 12Linq-Embed-MistralLinq AI Research
Older (2024) open 7B model that still ranks well; newer picks above usually beat it
Best for Free self-hosted search with an open 7B modelYou give up Needs an estimated 14 GB, more than higher-ranked small models like BidirLM
Not scored yetFree - 13multilingual-e5-large-instructMicrosoft
Small (560M) long-standing open choice for many languages that runs on modest hardware
Best for Multilingual search on laptops or small servers with about 2 GB of memoryYou give up Ranks below jina-embeddings-v5-text-small at the same memory size; jina is non-commercial
Not scored yetFree - 14GritLM-7BGritLM (Contextual AI)
Open 2024 model that both writes text and makes embeddings; Apache licence
Best for Small projects wanting one self-hosted model for both search and text repliesYou give up Ranks 34th on the board; needs an estimated 14 GB of memory
Not scored yetFree - 15e5-mistral-7b-instructMicrosoft
Older open 7B model; most newer open picks above rank higher at similar size
Best for Free self-hosted multilingual search if you can spare about 14 GB of memoryYou give up Ranks 36th; Qwen3-Embedding-8B ranks 4th with a similar memory need
Not scored yetFree - 16SFR-Embedding-2_RSalesforce
Salesforce's 2024 open 7B model; weights are licensed for non-commercial use only
Best for Research and other non-commercial search work on your own hardwareYou give up Licence bars commercial use; needs an estimated 14 GB and ranks 41st
Not scored yetFree - 17text-embedding-3-largeOpenAI
Easy hosted default if you already use OpenAI, but behind the open models above on this board
Best for Hosted embeddings at a published low price of $0.13 per million tokensYou give up Ranks 44th, below many free open models and gemini-embedding-001
Not scored yetPay as you go
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | harrier-oss-v1-27bMicrosoft · Open weights | Not scored | Ranked #1 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 1 of 470, mean 74.27. MTEB Multilingual v2 rank 1 -> #1 of distinct models; no complete independent second eval (RTEB beta lacks most top models) so no change Report a mistake | Free (open weights) | Teams that want the top open search model and can host it themselves | Needs an estimated 54 GB of memory at full size; plan on a data-center GPU |
| 2 | KaLM-Embedding-Gemma3-12BTencent · Open weights | Not scored | Ranked #2 of 466 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 2 of 466, mean 72.32. MTEB Multilingual v2 rank 2 -> #2 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Multilingual search on your own servers, a step down in size from the top pick | Needs an estimated 24 GB of memory; scores below harrier-oss-v1-27b |
| 3 | llama-embed-nemotron-8bNVIDIA · Open weights | Not scored | Ranked #3 of 466 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 3 of 466, mean 69.46. MTEB Multilingual v2 rank 3 -> #3 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Multilingual search on a single mid-size GPU, ranked third on the board | Licence may rule out commercial use; needs an estimated 16 GB of memory |
| 4 | Qwen3-Embedding-8BAlibaba · Open weights | Not scored | Ranked #4 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 4 of 470, mean 70.58. MTEB Multilingual v2 rank 4 -> #4 of distinct models (no rank spread shown); no complete independent second eval so no change Report a mistake | Free (open weights) | Commercial search projects that want a free model under a permissive licence | Needs an estimated 16 GB at full size; the smaller sizes rank lower |
| 5 | gemini-embedding-001Google · Closed | Not scored | Ranked #5 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 5 of 470, mean 68.37. MTEB Multilingual v2 rank 5 -> #5 of distinct models (no rank spread shown); no complete independent second eval so no change Report a mistake | Not confirmed | Teams that want hosted text embeddings rather than running their own model | Price not in our files; Google plans to shut it down on 14 May 2028 |
| 6 | Octen-Embedding-8BOcten · Open weights | Not scored | Ranked #7 of 461 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 7 of 461, mean 67.84. MTEB Multilingual v2 rank 7 -> #6 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | A free, permissively licensed alternative to Qwen3-Embedding-8B for search | Ranks below the Qwen3 model it is built on; needs an estimated 16 GB of memory |
| 7 | F2LLM-v2-14BAnt Group · Open weights | Not scored | Ranked #8 of 462 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 8 of 462, mean 68.74. MTEB Multilingual v2 rank 8 -> #7 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Teams that want one model family in several sizes to match their hardware | The 14B needs an estimated 28 GB of memory; the lighter sizes rank lower |
| 8 | Seed1.6-embeddingByteDance · Closed | Not scored | Ranked #11 of 461 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 11 of 461, mean 70.26. MTEB Multilingual v2 rank 11 (1215 version) -> #8 of distinct models; no complete independent second eval so no change Report a mistake | Not confirmed | Hosted multilingual search without running your own servers | Price not in our files; ranks below several free open models on the board |
| 9 | Giga-Embeddings-instruct-10Bai-sage (Sber) · Open weights | Not scored | Ranked #13 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 13 of 470, mean 65.64. MTEB Multilingual v2 rank 13 -> #9 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Teams hosting their own search who want less compute per request | Still needs an estimated 20 GB of memory to load all 10B parameters |
| 10 | jina-embeddings-v5-text-smallJina AI (Elastic) · Open weights | Not scored | Ranked #15 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 15 of 470. MTEB Multilingual v2 rank 15 (omni-small variant 14) -> #10 of distinct models; no complete independent second eval so no change Report a mistake | Free weights, non-commercial licence | Research and personal projects needing a small multilingual model on modest hardware | Free weights are non-commercial only; business use means paying for its API |
| 11 | BidirLM-Omni-2.5B-EmbeddingBidirLM · Open weights | Not scored | Ranked #17 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 17 of 470, mean 63.50. MTEB Multilingual v2 rank 17 -> #11 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Free self-hosted search on machines with around 5 GB of spare memory | Ranks 17th on the board, behind the larger open models above it |
| 12 | Linq-Embed-MistralLinq AI Research · Open weights | Not scored | Ranked #25 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 25 of 470. MTEB Multilingual v2 rank 25 -> #12 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Free self-hosted search with an open 7B model | Needs an estimated 14 GB, more than higher-ranked small models like BidirLM |
| 13 | multilingual-e5-large-instructMicrosoft · Open weights | Not scored | Ranked #27 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 27 of 470, mean 63.22. MTEB Multilingual v2 rank 27 -> #13 or #14 of distinct models either way; no complete independent second eval so no change Report a mistake | Free (open weights) | Multilingual search on laptops or small servers with about 2 GB of memory | Ranks below jina-embeddings-v5-text-small at the same memory size; jina is non-commercial |
| 14 | GritLM-7BGritLM (Contextual AI) · Open weights | Not scored | Ranked #34 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 34 of 470, mean 60.92. MTEB Multilingual v2 rank 34 -> #14 to #16 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Small projects wanting one self-hosted model for both search and text replies | Ranks 34th on the board; needs an estimated 14 GB of memory |
| 15 | e5-mistral-7b-instructMicrosoft · Open weights | Not scored | Ranked #36 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 36 of 470, mean 60.03. MTEB Multilingual v2 rank 36 -> #15 to #17 of distinct models; no complete independent second eval so no change Report a mistake | Free (open weights) | Free self-hosted multilingual search if you can spare about 14 GB of memory | Ranks 36th; Qwen3-Embedding-8B ranks 4th with a similar memory need |
| 16 | SFR-Embedding-2_RSalesforce · Open weights | Not scored | Ranked #41 of 470 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 41 of 470, mean 59.80. MTEB Multilingual v2 rank 41 -> #16 to #19 of distinct models; no complete independent second eval so no change Report a mistake | Free weights, non-commercial licence | Research and other non-commercial search work on your own hardware | Licence bars commercial use; needs an estimated 14 GB and ranks 41st |
| 17 | text-embedding-3-largeOpenAI · Closed | Not scored | Ranked #44 of 461 in a multilingual search-quality test (MTEB)NotesTheir figure: MTEB(Multilingual, v2) rank 44 of 461. MTEB Multilingual v2 rank 44 -> #17 to #20 of distinct models; no complete independent second eval so no change Report a mistake | API $0.13 per 1M tokens | Hosted embeddings at a published low price of $0.13 per million tokens | Ranks 44th, below many free open models and gemini-embedding-001 |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Ordered by the MTEB leaderboard, an independent test of search and retrieval quality. One row per model line, newest version only.
How we score this list. No score yet: we have not set a rule for turning MTEB results into a score. The order is the leaderboard's. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 18 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.