The best embedding models for RAG in 2026, ranked on dated evidence

The best embedding model for RAG in 2026 turns on one axis most lists bury. Can the weights run on your own hardware, or not? NVIDIA's Nemotron 3 Embed 8B, released 16 July 2026, leads the retrieval-focused RTEB board at 78.5 percent. Qwen3-Embedding-8B has held 70.58 on multilingual MTEB since June 2025. Gemini Embedding 2 posts 69.9 on its publisher's page. It also reads images, video, audio, and documents. This page ranks ten models on scores read at their sources on 28 October 2026. The ranking rules come first, then the table. Every row pairs with its local or hosted reality. This platform serves two rows, bge-m3 and wemm-embed. Both sit inside the field at the position their evidence supports.

Every score below names its owner and its date. Publisher claims come from the primary pages below, all read 28 October 2026:

• The Qwen3 Embedding card
• The Nemotron 3 Embed release post
• Google's Gemini Embedding page
• OpenAI's embeddings guide
• The Cohere Embed 5 post
• The bge-m3, WeMM, and harrier cards

Board positions come from the MTEB leaderboard site, read the same day. Third-party runs are labeled third-party. Where a number does not exist, the cell says so in words. Ranking pages today paste scores with no dates and no owners. One bills an Alibaba model under a name that reads as OpenAI's next. This page keeps dates, owners, and existence straight.

Last verified: 2026-10-28

The ranking rules, stated first

A ranked list without stated rules is a mood. These are the rules, in order of weight.

1. Retrieval evidence before generic averages. RAG lives or dies on retrieval. RTEB, a benchmark of 30 retrieval tasks across legal, finance, code, and healthcare in 22 languages, per its own board description, counts more than a general average that mixes in clustering and sentiment tasks.
2. Named and dated scores only. Every number names the party that measured it and the date it was measured or read. A score with no owner is not evidence.
3. Same-day reads. Every board position and publisher score on this page was read on 28 October 2026, the day this page was written. The boards move weekly.
4. Deployment reality as a first-class axis. Open weights change the cost and the exit. A model you can run, move, and keep is not the same product as an endpoint you rent.
5. Dimensions and windows as secondary facts. They matter, and the full per-model table with truncation behavior lives in the dimensions and context guide. This page carries them as columns, not as the ranking.
6. Cost as a shape, not a pasted number. Which lane a model sells through, open or hosted or both. The dated rate comparisons live in the embedding pricing guide.

One more rule governs the whole page. Publisher tables promote the publisher's model. That is how release tables work. Where a claim is the publisher's own, the row says so.

The boards, read today

Three public boards answer three different questions. All three were read 28 October 2026 on the MTEB leaderboard site.

Three facts hang off that table. First, the RTEB board's top row today is not a product. It is a voyage-4-large run with evolved prompts, submitted by MongoDB, Voyage's parent company, in a 2,048-dimension configuration. The plain voyage-4-large row sits sixth. The best model you can simply download and run on that board is NVIDIA's, at rank two. Second, the general board and the retrieval board disagree about who leads. That disagreement is the most useful thing a buyer can know about embedding benchmarks in 2026. Third, the English board's size-group winners span an 8B model, a 4B model, and a sub-500M model. The 27B leader lives on the multilingual board, not the English one. The word best always needs a size qualifier attached.

The ranked table

Ten rows, ranked on the six rules above. Read the evidence column as the load-bearing one.

Two honest notes on the columns. The dims and window cells repeat facts this site already owns in full. The model-by-model table carries sixteen rows, with truncation stops and per-provider caveats. This table only points at it. The weights column states what the publisher's own pages state. Where a public score does not exist, the row says so plainly. Cohere prints no dated public benchmark number on its Embed 5 post. Its row ranks on capability facts, not a score.

Row by row, the evidence

1. Nemotron 3 Embed: the July 2026 event

NVIDIA released Nemotron 3 Embed on 16 July 2026, as three open models. The 8B is the flagship. A 1B efficiency model and a 4-bit hardware-optimized 1B variant ship beside it. Its release post states the claim plainly. On RTEB, the 8B ranks first, scoring 78.5 percent on RTEB and 75.5 percent on multilingual MTEB retrieval. Publisher and board agree within one row: today's board read puts the 8B second, behind the tuned MongoDB submission. What RAG teams ask for shows up in the same post. A 32k window covers long documents and multi-turn agent histories. Code retrieval, open training data, and day-one support in local serving stacks round it out. Listed specs: 4,096 dimensions, roughly 8B parameters, 34 languages. For a retrieval-first RAG decision in late 2026, this family has the freshest dated evidence and no vendor lock.

2 and 5. Qwen3 Embedding: the proven open flagship, two sizes

Qwen's Qwen3 Embedding card carries the claim the whole shelf repeats. Its 8B model ranked first on the MTEB multilingual leaderboard as of 5 June 2025, at 70.58. Today's board read still shows 70.58 for the 8B, now at rank four under two newer families. Card tables add the rest. Multilingual, the 4B scores 69.45. On English v2, the 8B scores 75.22. Windows run 32k, languages cover more than 100, and queries take instructions. Output dimensions are user-defined under Matryoshka training. Both sizes ship Apache 2.0. A third-party roundup updated 23 October 2026 adds the retrieval-specific cut. On a June 2026 BEIR run, Qwen3-Embedding-8B scored 56.4 nDCG@10 against bge-m3's 51.8. Quiet value row: the 4B, which leads its size group on both boards read today, at 69.45.

3. Gemini Embedding 2: the hosted all-rounder

Google's Gemini Embedding page posts 69.9 on multilingual MTEB for the new model. Previous generation: 68.4. What no text-only row in this table has also shows up on that page. Audio retrieval scores 73.9 on the publisher's speech benchmark. Image, video, and document inputs ride the same model. Window size is 8,192 tokens across input types. Output runs 128 to 3,072 dimensions with trained truncation, and the page states coverage over 100 languages. One honesty note follows. On the public multilingual board read today, the Gemini Embedding 2 row is a preview entry with no complete mean. Its older sibling, gemini-embedding-001, carries 68.37 at rank five. So 69.9 stands as the publisher's number. Teams wanting one hosted model for mixed text, image, and audio retrieval have their row.

4. voyage-4-large: retrieval board leader, closed weights

Voyage shipped the voyage-4 family on 15 January 2026. Its own docs list voyage-4-large at 32,000 tokens of context and 1,024 default dimensions, and call it the best general-purpose and multilingual retrieval quality in the catalog. On RTEB, the board read today gives this family the strongest neutral evidence any closed model has. Rank one: the tuned 2,048-dimension submission. Sixth: the plain row. That same family stays off the top of the general multilingual board, where its rows carry no complete mean. Trade: the weights. Closed means one vendor and no local exit. If the endpoint ever changes price or policy, the whole index re-embeds at whatever the next vendor charges.

6. harrier-oss-v1-27b: the score ceiling, at a price

harrier-oss-v1-27b tops the general multilingual board read today at 74.27. It ships MIT, in three sizes: 270m, 0.6b, and 27b. Card tables put the small ones at 66.5 and 69.0. This 27B carries 5,376 dimensions and a 32,768-token window per its card. Here is the raw score ceiling of the open field today. Cost is literal: a 27B encoder needs serious hardware. On the board listing, memory sits at about 50 GB for this model, against about 15 GB for the 8B-class rows beside it. For most RAG teams the 0.6b sibling at 69.0 is the usable row. Cite the 27B when someone claims a small model leads everything.

7. bge-m3: the 2024 workhorse this platform serves

bge-m3's card still describes a shape no 2026 flagship replaced. One pass returns dense, sparse, and multi-vector outputs. It works across more than 100 languages, over 8,192 tokens, under MIT weights. For RAG, the card's own advice is hybrid retrieval plus reranking. Use the dense and sparse outputs together, then filter with a reranker. Its 2026 numbers are honest about its rank. A third-party roundup updated 23 October 2026 reads it at 59.56 on multilingual MTEB. A June 2026 BEIR run gives it 51.8 nDCG@10. On the public board today it places 52nd of 483, with no complete mean posted. Throughput and control keep it on production stacks anyway. That same third-party benchmark measured about 4,800 passages per second at batch 64 on consumer Apple Silicon, roughly eight times the Qwen3-Embedding-8B rate. Memory footprint stays small, and the license never asks again.

8. text-embedding-3: the incumbent default

OpenAI's embeddings guide still lists three rows: small, large, and the legacy ada. The big two read 62.3 and 64.6 percent MTEB, with 8,192-token inputs. Those numbers are unchanged since the 2024 launch. No new OpenAI embedding model has shipped since. The pair remains the pragmatic default for teams already inside that ecosystem. One endpoint, trained dimension shortening, no hardware. The model-level comparison against bge-m3, with independent May 2026 retrieval runs, lives in the bge-m3 versus text-embedding-3 guide.

9. Embed 5: enterprise retrieval, 128k window

Cohere's Embed 5 post ships two tiers, Pro and Fast, over one shared vector space. Index with Pro, query with the cheaper Fast, no re-embed. The feature set is the draw. Text and image inputs, more than 100 languages, and a 128k-token window lead it. Six trained dimension stops run from 2,048 down to 256. Int8 and binary outputs ship too, and the post states a self-hosting option. The post calls Pro's retrieval performance its strongest to date, but prints no dated public number. So this page ranks the row on capability, not score.

10. WeMM-Embedding-2B: the multimodal RAG row, served here

Tencent's WeMM card tops the open 2B rows on MMEB-v2 at 77.9. Qwen3-VL-Embedding 2B, the baseline, reads 73.2. One space covers text, images, video, visual documents, and interleaved mixes. Reads run up to 262,144 positions. This is the row for RAG over scans, screenshots, slide decks, and video, the corpora where text-only models see nothing. Audio is not supported, and the card says so itself. Larger siblings score higher still, 79.2 at 4B and 80.6 at 9B. This platform serves the 2B.

What the ranking pages get wrong

Four corrections, all checked today against primary sources.

A name that reads as OpenAI's next move. One ranking page bills text-embedding-v4 in its headline cost table, at an undated per-million rate, right beside the OpenAI rows. The model is real, and it is not OpenAI's. The incumbent's own table attributes the row to Alibaba's Model Studio. OpenAI's own embeddings guide, read 28 October 2026, lists no such model. It lists three: 3-small, 3-large, and ada-002. An unwary reader takes the name as OpenAI's next generation, at a price with no date. A cost table that mixes owners without labels is the loudest possible argument for the owner-and-date rule.

Vanished pages. Two of the four ranking pages captured in this research wave's search logs on 26 October returned page-not-found on 28 October, two days later, both relocated to new URLs by their owners. The lane churns that fast. An undated listicle is already drifting the day it ships.

Stale scores drift quietly. One incumbent prints 68.32 for Google's previous-generation embedding model. The board read today reads 68.37 for the same row, and Google's own page rounds it to 68.4. None of the three numbers is wrong on its own date. Only the date separates a fact from a fabrication, which is why every cell here carries one.

The average hides the task you care about. A general MTEB mean blends retrieval with clustering, classification, and pair tasks. RAG needs the retrieval slice. The board disagreement documented above, voyage leading retrieval while harrier leads the general mean, is the pattern in the wild. Run a small retrieval evaluation on your own corpus before committing the index. Every methodology-aware source read for this page says the same sentence.

Local weights or a hosted route, per model

This is the axis the 2026 listicles skip. It decides multi-year projects.

One rule falls out of the table. Open weights buy you an exit. Hosted routes buy you operations. The crossover math has a worked answer. At some monthly token volume, your own GPU beats any per-token rate. The embeddings break-even guide works that math to a number. One more lane sits between the two. A third-party model marketplace resells many providers' embedding rows behind one endpoint, open weights included, at deep discounts. Its rates and catalog are mapped in the pricing guide, which owns that surface.

Where this platform's rows sit

The embeddings route here serves two rows as of 28 October 2026. This section states their position without inflating it. Neither row tops the 2026 text-retrieval boards. Both hold their own corners.

bge-m3 is the text row. It fixes 1,024 dimensions, an 8,192-token window, more than 100 languages, hybrid dense-sparse-multi-vector output, MIT weights. Planning numbers from the platform's 28 August 2026 bench window: median latency 1.38 seconds single-flight, 1.90 at eight in flight, 4.33 at thirty-two. Aggregate throughput scales about ten times along that curve. One real request completed the full round trip on 9 September 2026. Its usage row matched the billing ledger. Read them as planning inputs from those windows, not promises.

wemm-embed is the multimodal row. It carries 2,048 dimensions over text, images, video, and visual documents, with 262,144 positions and six trained truncation stops. Its measured lane throughput, 2026-09-25: 136 tokens per second at peak, eight requests in flight.

Rate position, stated as relations and dated. Closed flagships sit above this platform's rate. Discount hosts serving the same open weights sit below it. The full dated comparison, every provider, batch tiers included, is the pricing guide's job. Live rates for these two rows sit on the pricing page.

What the lane sells is the aggregate. Two rows, one request shape, one prepaid wallet. Open weights sit underneath both, if you ever want to take the corpus home. The young marketplace class compresses rates at the bottom of the market. It does not hand you a hybrid retrieval model and a multimodal model on one bill.

Make the first call

Sign up, create a key, and embed two chunks:

curl -X POST "https://api.ironstratum.com/v1/embeddings" \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-m3",
    "input": [
      "One chunk of the corpus, one thousand twenty four numbers back.",
      "A second chunk, metered per token against the wallet."
    ],
    "encoding_format": "float"
  }'

Each vector returns 1,024 floats, and the usage block in the response is the metering record. A refused request bills nothing. The bge-m3 model page holds the full request contract, and the models index is the live catalog.

Pick by corpus, in 60 seconds

Work the list and stop at the first fit.

1. Retrieval quality is the product, weights must be yours: Nemotron-3-Embed-8B, then the Qwen3 4B if hardware is tight.
2. One hosted model, mixed text, image, audio, and document queries: Gemini Embedding 2.
3. English-first enterprise docs, 128k windows, image pages, and a query-price tier: Cohere Embed 5.
4. Already inside the OpenAI ecosystem and happy there: text-embedding-3, read the head-to-head before you commit.
5. The corpus is multilingual, or queries cross language lines, and you want hybrid keyword-plus-vector retrieval from one model: bge-m3.
6. The corpus is scans, screenshots, decks, or video: wemm-embed, or its bigger siblings self-hosted.
7. You have serious hardware and want the board ceiling: harrier-oss-v1-27b.
8. Storage is the constraint more than quality: any Matryoshka-trained row above, cut to 256 dimensions; the dimensions guide owns that dial.
9. Retrieval still misses after a good embedder: add a reranker. The two-stage decision and the cross-encoder explainer own that step.

Common questions

What is the best embedding model for RAG in 2026? On retrieval-dated evidence: Nemotron-3-Embed-8B for open weights, Gemini Embedding 2 for a single hosted multimodal model, voyage-4-large where a closed pure-retrieval leader fits your stack. The honest full sentence: the best model is the one whose evidence class, deployment lane, and window fit your corpus, and the table above is how you read that.

Is MTEB score enough to pick a model? No. The MTEB average blends retrieval with clustering and classification tasks, and the boards disagree at the top: the leader of the general multilingual board and the leader of the retrieval board are different models, both read 28 October 2026. Use retrieval-specific scores where they exist, then run a small evaluation on your own corpus before building the index.

What is the best free or local embedding model for RAG? Free means weights. Qwen3-Embedding (0.6B to 8B, Apache 2.0), bge-m3 (MIT), harrier-oss-v1 (MIT, three sizes), and Nemotron 3 Embed (open, commercial use stated) all run on hardware you control with no per-token bill. The machine, power, and operator time are the real bill, and the break-even guide works that math to a crossover point.

Do more dimensions mean better RAG? No. Width is storage cost, not accuracy. Publisher tables show wide models cut to a quarter of their size still beating older full-width models, and the measured ladders live in the dimensions guide. Choose quality by retrieval evidence, then choose width by your storage budget.

Do I need a reranker if the embedding model is good? Usually, yes, for precision at the top of the list. The bge-m3 card itself recommends hybrid retrieval followed by reranking for RAG pipelines. Reranking costs a second model and a second latency budget; the two-stage retrieval decision covers when it pays.

Which rows does this platform serve? Two, as of 28 October 2026: bge-m3 for multilingual hybrid text retrieval and wemm-embed for multimodal work, both metered per token against one prepaid wallet. The pricing page carries the live rates.

The short version

Rank on retrieval evidence, read the day you decide. Nemotron 3 Embed 8B owns the freshest retrieval claim at 78.5 percent RTEB, released 16 July 2026. Qwen3-Embedding-8B has held 70.58 multilingual MTEB since June 2025. Gemini Embedding 2 posts 69.9 with every input type in one model. The boards disagree at the top, so name the benchmark or the ranking is noise. Open weights change the exit more than any score column does. This platform serves bge-m3 and wemm-embed, placed mid-field on the evidence above, one bill, both open-weight. When the question turns to width, windows, rates, or the rerank stage, the dimensions table, the pricing guide, and the two-stage guide own those answers. Ready to measure on your own corpus? Sign up and embed two chunks.

Last verified: 2026-10-28