Embedding Model Leaderboard 2026: MTEB Rankings, RAG-Specific Recall, and the Open vs API Trade-off

The Massive Text Embedding Benchmark is the closest thing the embedding world has to a standardized exam. Maintained by Hugging Face, MTEB v2 covers over 50 tasks across retrieval, classification, clustering, semantic textual similarity, summarization, and reranking, drawing on datasets from multiple languages and domains. A model's MTEB average score aggregates performance across all of these, giving a single headline number that lets you compare models without running your own experiments on every task type.

The benchmark's strength is also its weakness. Because the average pools more than 50 heterogeneous tasks, a model can score well by excelling at classification and clustering while being mediocre at the dense retrieval task that powers most RAG pipelines. The inverse is also true: a model tuned specifically for passage retrieval may rank lower on the overall leaderboard while outperforming the headline winner on the exact task you care about. This is why the MTEB-Retrieval subset deserves separate attention, which the next section covers in detail.

There are additional limits worth naming. MTEB primarily evaluates English-language performance, though the MMTEB multilingual extension covers additional languages for some models. It tests on held-out academic datasets, which may not reflect your domain's vocabulary, query style, or document length distribution. And it evaluates embedding quality in isolation — not how that quality translates into end-to-end RAG answer quality when combined with a specific chunking strategy and retriever. For a more practical grounding, pairing MTEB scores with your own corpus evaluation (described in the steps section below) is strongly recommended.

NV-Embed-v2 from NVIDIA holds the top position on MTEB v2 average at approximately 72.3 as of June 2026. This is a 7.85-billion-parameter model that leverages instruction-following capabilities derived from the NV-Embed architecture, producing 4096-dimensional vectors with a 32,768-token context window. Its strength lies in following task-specific instructions at inference time, allowing the same model weights to serve retrieval, classification, and semantic similarity tasks with differentiated prompting. The trade-off is significant GPU memory requirement — an A100 or equivalent is the practical minimum for production inference at reasonable throughput.

Voyage AI's voyage-3-large sits in second at approximately 70.1 MTEB average, but as the next section explains, it leads on the MTEB-Retrieval subset at roughly 62–63 — meaning it is the better choice for most RAG applications despite its lower headline ranking. Alibaba's gte-Qwen2-7B-instruct takes third at approximately 69.5, benefiting from the Qwen2 language model backbone to produce rich 3584-dimensional embeddings. Cohere's embed-v4.0 ranks fourth at approximately 68.2, with the notable differentiator of native 128,000-token context — a meaningful advantage for embedding long documents without pre-chunking.

The remainder of the top ten — bge-multilingual-gemma2, voyage-3, mxbai-embed-large-v1, text-embedding-3-large, jina-embeddings-v3, and text-embedding-3-small — span a range from roughly 67 down to 62. Within this band, the differences in practical retrieval quality are often smaller than the differences implied by the absolute numbers, and operational factors like latency, API reliability, token limits, and total cost often outweigh a two-point score gap. The embeddings cost calculator can help model the cost side of this comparison.

MTEB-Retrieval is the subset of MTEB tasks focused specifically on dense passage retrieval — the task of finding the most semantically relevant documents in a corpus given a natural-language query. This is the core competency for RAG pipelines, vector database search, and semantic search applications. The reordering it produces compared to the overall average is significant and worth understanding before committing to any model.

voyage-3-large leads the retrieval subset at approximately 62–63, climbing above NV-Embed-v2 (which scores around 61 on retrieval despite leading the overall average). Cohere embed-v4.0 follows closely at roughly 60–61. gte-Qwen2-7B-instruct drops more substantially to approximately 59, suggesting that its broad capability advantages across classification and clustering tasks account for more of its headline score than its retrieval performance. OpenAI's text-embedding-3-large, despite ranking eighth overall, scores approximately 58–59 on retrieval, narrowing the gap with models that rank significantly above it on the aggregate.

The practical takeaway is this: if you are building a RAG system, the MTEB-Retrieval column is the one to sort by, not the MTEB average. The divergence between the two rankings is not noise — it reflects genuine differences in how different model architectures and training objectives handle the query-document matching problem. Models that excel at classification tasks use their embedding space differently than models optimized for retrieval, and the aggregate average obscures this. For RAG-specific guidance on how the embedding layer interacts with retrieval architecture choices, see the RAG architecture decision tree 2026 and the GraphRAG vs Vector RAG comparison.

NV-Embed-v2 is the current open-source benchmark leader and the model to beat for teams with GPU infrastructure. Released under Apache 2.0, it is freely usable for commercial applications without per-token fees. The model's instruction-following architecture means you prompt it differently depending on the task — for retrieval, the query should be prefixed with a retrieval-specific instruction string, while document embeddings are typically computed without instruction prefixing. This asymmetric prompting is well-documented in NVIDIA's model card on Hugging Face. At 7.85 billion parameters and 4096-dimensional output, it is memory-intensive: expect approximately 16GB of GPU VRAM at half-precision, making an A100 40GB or H100 the realistic minimum for production batch inference. Quantized versions are available and reduce memory requirements, with modest quality trade-offs.

gte-Qwen2-7B-instruct from Alibaba ranks third overall and benefits from the Qwen2-7B language model backbone, which was pre-trained on a large multilingual corpus before embedding-specific fine-tuning. The 3584-dimensional output is dense and requires proportionally larger vector indices, which has downstream effects on both memory and query latency in production. Its 32,768-token maximum context is a practical advantage for embedding longer documents without chunking, though embedding quality at the extreme end of the context window degrades, as with most long-context models. Like NV-Embed-v2, it is Apache 2.0 licensed and requires self-hosted GPU infrastructure.

bge-multilingual-gemma2 from BAAI occupies a different niche. While its 67.4 MTEB average is strong and its retrieval score of approximately 56 is competitive, its primary advantage is multilingual coverage. Built on the Gemma2 backbone, it supports dozens of languages with notably stronger cross-lingual retrieval performance than English-focused models. However, its 8,192-token context window is shorter than the other open-source leaders, and its 3584-dimensional output carries the same index size implications as gte-Qwen2. Teams building multilingual RAG applications or semantic search over non-English corpora should benchmark this model alongside Cohere's embed-v4.0, which offers similar multilingual breadth via API.

OpenAI's text-embedding-3-large remains the most widely deployed embedding model in production systems as of mid-2026, reflecting its combination of reasonable quality, ecosystem integration, and operator familiarity. At approximately 64.6 MTEB average and 58–59 on retrieval, it is not the top performer in either ranking — but it benefits from extremely high uptime, predictable latency, and native integration with LangChain, LlamaIndex, and most major RAG frameworks. The 3072-dimensional output is large enough to represent nuanced semantic distinctions, and Matryoshka representation learning means you can truncate to smaller dimensions (1536, 768, etc.) with modest quality loss if index storage is a constraint. OpenAI's OpenAI embeddings rate limits are relevant at scale and worth reviewing before designing your ingestion pipeline. Pricing at approximately $0.13/1M tokens is moderate — not the cheapest option, but competitive with the quality level it delivers.

Voyage AI's voyage-3-large is the strongest API option for pure retrieval performance, leading the MTEB-Retrieval subset at approximately 62–63 despite ranking second on the overall average. Voyage AI was acquired by MongoDB in 2025 and its models are now available both through the Voyage API and natively within MongoDB Atlas Vector Search. The 2048-dimensional output is smaller than NV-Embed-v2 or gte-Qwen2, which reduces index storage requirements while maintaining strong retrieval quality — a favorable trade-off. At approximately $0.18/1M tokens, it is the most expensive API option on the list, but the retrieval quality premium is measurable. voyage-3 at approximately $0.06/1M is a lower-cost alternative from the same provider with somewhat lower retrieval performance, suitable for applications where cost efficiency outweighs peak performance. For a detailed comparison, see Cohere vs Voyage vs OpenAI embeddings.

Cohere's embed-v4.0 offers the most distinctive feature profile of any API model: a 128,000-token context window, which is an order of magnitude larger than most competitors, and leading multilingual performance on the MMTEB benchmark across more than 100 languages. The 1024-dimensional output is smaller than the open-source leaders, which helps with index size and query speed. At approximately $0.10/1M tokens, it is mid-range on price. The long-context capability is particularly useful for embedding entire documents rather than chunked passages — though this changes the chunking strategy decisions described in chunking strategies 2026. Jina AI's jina-embeddings-v3 rounds out the API category at approximately $0.02/1M tokens, the joint-lowest API price on the list alongside text-embedding-3-small. Its CC BY-NC 4.0 license means it is free for non-commercial self-hosting but requires a commercial license or API payment for production commercial use.

MTEB measures performance across general-domain academic datasets, and the rankings above are meaningful for typical text corpora — web content, news, product descriptions, conversational data. But for highly specialized corpora, domain-specific embedding models trained on in-domain text can outperform general-purpose leaders by five to ten points on the relevant retrieval task. This gap is large enough to materially affect answer quality in production RAG systems.

voyage-code-3 is the clearest example. On the CodeSearchNet benchmark, which tests code-to-code and natural-language-to-code retrieval, voyage-code-3 outperforms general models including its sibling voyage-3-large by a significant margin. The model is trained on a large corpus of code in multiple programming languages and understands function signatures, variable names, and code documentation patterns in ways that general models do not. For any RAG application over a code repository, documentation codebase, or technical knowledge base with substantial code content, benchmarking voyage-code-3 before committing to a general model is strongly advised.

Voyage AI also maintains voyage-finance-2 for financial documents and voyage-law-2 for legal corpora, both of which show meaningful performance improvements over general models on their respective domain benchmarks. Cohere's embed-v4.0, while a general-purpose model, leads on multilingual benchmarks and is the recommended starting point for any application requiring embedding quality across non-English languages — its MMTEB scores reflect consistent multilingual retrieval performance that none of the open-source models on this list currently match at comparable operational cost. When evaluating whether a domain-specialized model is justified, the practical test is to build a 100–500 pair evaluation set from your actual corpus and compare recall@10 directly, rather than relying on public benchmark scores as a proxy.

The make-vs-buy decision for embedding models comes down to a total cost of ownership calculation that most teams do not run carefully enough before committing. The API cost is visible and linear: at OpenAI's text-embedding-3-large rate of approximately $0.13/1M tokens, embedding 100 million tokens per month costs $13. At NV-Embed-v2 self-hosted, the per-token fee is zero, but the infrastructure cost is not. A single A100 80GB instance on a major cloud GPU provider runs approximately $2–4 per hour, translating to roughly $1,500–$3,000 per month for a single GPU at continuous operation. The break-even point between self-hosting NV-Embed-v2 and using text-embedding-3-large at that pricing is therefore somewhere above 10 billion tokens per month — a scale most applications do not reach.

The math shifts when you compare against more expensive API options. voyage-3-large at $0.18/1M tokens reaches cost parity with a single A100 at around 8–17 billion tokens per month depending on the GPU cost and utilization rate. At 50 million tokens per month, API wins decisively on cost for any provider on this list. At 500 million tokens per month, the comparison becomes sensitive to the specific GPU instance type, utilization rate, and whether inference can be batched efficiently. Teams already operating GPU clusters for LLM inference can attach an embedding model to existing infrastructure at near-zero marginal cost, which changes the calculation entirely.

There are non-cost factors that favor API models independent of the token volume calculation. API providers handle model versioning, scaling, uptime, and inference optimization. Self-hosted models require engineering time for deployment, monitoring, scaling policy design, and ongoing maintenance — costs that are real but rarely appear in a back-of-envelope TCO calculation. On the other side, self-hosted models give you data residency control, no rate limit exposure, and the ability to fine-tune on proprietary data. The embeddings cost calculator can help model the cost side at your specific token volumes; add 20–40% to any self-hosted estimate to account for engineering overhead and operational margin.

MTEB scores are a useful prior, but the only reliable way to know which embedding model will perform best on your corpus is to measure directly. The process is straightforward and does not require large resources. Start by sampling 200–500 query-document pairs from your actual data — ideally a mix of queries that represent real user questions and the documents that should surface as top results. Include both easy positives (clearly relevant) and hard negatives (plausibly relevant but incorrect) to make the evaluation discriminating. Label relevance at a binary or three-tier level (relevant, partially relevant, not relevant).

For each model under evaluation, embed all queries and documents, then compute recall@k (typically k=5 and k=10) and mean reciprocal rank. These are the metrics that translate most directly into end-user experience in a RAG system — recall@10 tells you how often the correct document appears in the top 10 retrieved results, which is the effective recall ceiling for the downstream LLM. Run this evaluation for each candidate model and compare. If you are testing chunking strategies in parallel, this benchmark structure also gives you the infrastructure to measure chunking impact independently of model choice; the chunking strategies 2026 post covers this in detail.

Once you have a baseline, run the benchmark on a monthly cadence. The MTEB leaderboard releases new strong models frequently, and the open-source tier in particular has been moving quickly. A discipline of re-evaluating when a new model posts a 3+ point MTEB-Retrieval improvement over your current model prevents gradual performance decay as better options become available. Keep your evaluation set stable over time so improvements are attributable to the model rather than evaluation drift. If your corpus changes significantly — due to domain expansion or new document types — update the evaluation set accordingly to keep it representative.

Every score in this article is derived from the public MTEB v2 leaderboard hosted on Hugging Face at huggingface.co/spaces/mteb/leaderboard. The leaderboard accepts community submissions and is updated continuously, which means rankings can change between the date this was written and the date you are reading it. New model releases from major labs and from independent researchers regularly displace existing entries. The numbers here should be treated as a snapshot from approximately June 2026, not as permanent ground truth.

For API pricing, the approximate figures cited in the table above were sourced from provider pricing pages as of June 2026. Pricing changes with product updates, competitive pressure, and volume discounts. Before budgeting any embedding-dependent system, verify current pricing directly at the provider: platform.openai.com/docs/pricing for OpenAI, docs.voyageai.com/docs/embeddings for Voyage AI, cohere.com/pricing for Cohere, and jina.ai/embeddings for Jina AI. Volume pricing, committed-use discounts, and enterprise agreements can alter the per-token cost significantly from published list prices.

Model architecture details, dimension counts, and context window lengths are sourced from the respective model cards on Hugging Face and provider documentation pages. Dimension counts reflect default output; models supporting Matryoshka Representation Learning can produce smaller vectors by truncating the full output, with quality degrading approximately proportional to the compression ratio. The Apache 2.0 licenses on the open-source models in this list are accurate as of the publication date but should be re-verified before commercial deployment — license terms can be updated with new model versions.