Best Embedding Models 2026: Ranked by Quality and Price
The ranking (September 2026)
The rank is a recommendation, not a score ordering; the score column is Mean (Task) on MTEB(Multilingual, v2) where a model has complete coverage of all 131 tasks. Note the biggest caveat first: the 2026 commercial flagships are barely benchmarked. In MTEB’s public results repository, voyage-4, voyage-4-large and Cohere Embed v4 have results for 2 of 131 multilingual tasks and 0 of 38 English tasks, and Gemini Embedding 2 has none. Any “MTEB score” quoted for those models is not a benchmark result.
| # | Model | Maker | MMTEB v2 | Dims | Context | Price / 1M tokens | License | Best for |
|---|---|---|---|---|---|---|---|---|
| 1 | Qwen3-Embedding-8B | Alibaba | 70.75 | 4096 (MRL 32+) | 32,768 | Free (self-host) | Apache 2.0 | Best all-round |
| 2 | voyage-4-large | Voyage (MongoDB) | not published | 1024 default (256–2048) | 32,000 | $0.12 | Proprietary | Best commercial API |
| 3 | Gemini Embedding 2 | not published | 128–3072 | 8,192 | $0.20 text | Proprietary | Best multimodal | |
| 4 | harrier-oss-v1-27b | Microsoft | 74.31 | 5376 | 32,768 | Free (self-host) | MIT | Top of the multilingual board |
| 5 | Cohere Embed v4 | Cohere | not published | 256/512/1024/1536 | 128,000 | not published | Proprietary | Longest API context |
| 6 | Qwen3-Embedding-4B | Alibaba | 69.57 | 2560 | 32,768 | Free | Apache 2.0 | Best quality per GB |
| 7 | granite-embedding-311m-r2 | IBM | not published | 768 (MRL 128+) | 32,768 | Free | Apache 2.0 | Enterprise-clean licensing |
| 8 | text-embedding-3-large | OpenAI | incomplete (109/131) | 3072 | 8,192 | $0.13 | Proprietary | Widest tooling support |
| 9 | harrier-oss-v1-0.6b | Microsoft | 69.17 | 1024 | 32,768 | Free | MIT | Best small open model |
| 10 | Nemotron-3-Embed-8B | NVIDIA | incomplete | 4096 | 32,768 | Free | OpenMDW-1.1 | Best on RTEB |
| 11 | voyage-4-lite | Voyage | not published | 1024 default | 32,000 | $0.02 | Proprietary | Cheapest good API |
| 12 | EmbeddingGemma-300m | 61.15 (model card) | 768 (MRL 128+) | 2,048 | Free | Gemma Terms | On-device (<200 MB RAM) | |
| 13 | text-embedding-3-small | OpenAI | incomplete | 1536 | 8,192 | $0.02 | Proprietary | Cheapest OpenAI |
| 14 | bge-m3 | BAAI | not published | 1024 | 8,192 | Free | MIT | Legacy default (35M downloads/month) |
What changed since 2025
Open weights caught up, then the commercial vendors stopped publishing. The top of MTEB(Multilingual, v2) is now entirely open: Microsoft’s harrier-oss-v1-27b at 74.31, Tencent’s KaLM-Embedding-Gemma3-12B at 72.32, Qwen3-Embedding-8B at 70.75. The best closed model with complete coverage is gemini-embedding-001 at 68.84 — already superseded by Gemini Embedding 2, which Google did not submit.
Several 2025 defaults froze. BAAI has shipped no new general-purpose text embedder since July 2024; intfloat’s E5 family stopped in February 2024; OpenAI’s last embedding release was January 2024; Nomic published no 2026 embedding model. Qwen3-Embedding itself dates from June 5, 2025 with no text successor.
New entrants are not the usual suspects. Microsoft (Harrier, March 2026, MIT), NVIDIA (Nemotron-3-Embed-8B, July 16, 2026), IBM (Granite 311m r2, April 29, 2026) and Perplexity (pplx-embed-v1, February 2026) all shipped open-weight embedders.
The specialists got acquired. Voyage → MongoDB (February 2025), Jina → Elastic (October 2025), ZeroEntropy → Notion (July 24, 2026; hosted products sunset September 4, 2026, models relicensed Apache 2.0). Of the independents, only Cohere and Nomic remain.
The commercial APIs in depth
Voyage 4 — best commercial API, watch the data policy
The Voyage 4 family (January 15, 2026) is a mixture-of-experts line with one property nobody else matches: all four models share one embedding space. Embed documents with voyage-4-large and queries with voyage-4-nano, or upgrade your query model later, without re-vectorising. Every model supports 2048/1024/512/256 dimensions and five output precisions (float32, int8, uint8, binary, ubinary) at a 32K context. Prices: voyage-4-large $0.12, voyage-4 $0.06, voyage-4-lite $0.02 per million tokens, first 200M tokens free, 33% Batch discount. voyage-4-nano is open weights (Apache 2.0) — the only self-hostable model that stays space-compatible with a hosted frontier model. Specialists include voyage-code-4 (August 13, 2026, $0.12).
The catch: opting out of Voyage storing and training on your data requires a payment method on file and org-admin rights; free-allowance users cannot opt out, and opting out is one-way.
Gemini Embedding 2 — best multimodal, worst lock-in
Google’s first natively multimodal embedder: text, up to six images per request, video up to 120 seconds, native audio and PDFs up to six pages in one vector space. Text context is 8,192 tokens (4x gemini-embedding-001), with continuous Matryoshka dimensions 128–3072. Pricing: $0.20 per million text tokens, $0.45 image, $6.50 audio, $12.00 video. Google’s own docs are blunt: the embedding spaces of gemini-embedding-001 and gemini-embedding-2 are incompatible — upgrading means re-embedding everything, and Google’s deprecation history (text-embedding-004 got about six weeks’ notice before shutdown on January 14, 2026) means it will eventually force you to. Free-tier inputs may be used for training and reviewed by humans; EEA/UK/CH users get paid-tier treatment on free quota.
Cohere Embed v4 — longest context, oldest flagship
Shipped April 15, 2025 and still Cohere’s newest embedder as of September 2026. Its edge is 128,000 tokens of context — 15x OpenAI’s, 4x Voyage’s — returned as a single vector from interleaved text and images, at 256/512/1024/1536 dimensions with five precisions. It pairs naturally with Cohere Rerank in one vendor. Published per-token pricing is not available on the public page; check your contract.
OpenAI text-embedding-3 — ubiquitous, unrefreshed
No new model since January 25, 2024. text-embedding-3-large (3072 dims, 8,192 tokens) costs $0.13 per million ($0.065 via Batch — the “$0.065 price” seen online is the batch rate, not a cut); text-embedding-3-small costs $0.02. ada-002 ($0.10) is still live and not deprecated. Choose OpenAI for SDK ubiquity and Azure’s unusually stable deprecation schedule (all embedding deployments retire February 9, 2028), not for quality or multilingual coverage.
The open-weight picks
- Qwen3-Embedding-8B / 4B — Apache 2.0, 100+ languages, 32K context, MRL down to 32 dims, strongest off-the-shelf code retrieval on the independent CORE-Bench. Needs a real GPU.
- Microsoft harrier-oss-v1 (27B / 0.6B / 270M) — MIT, 32K context at every size; the 27B tops the multilingual board and the 0.6B is the best small open model at 69.17.
- IBM granite-embedding-311m-multilingual-r2 — Apache 2.0, 32K context, ~2,935 docs/s on one H100; the enterprise-legal-friendly choice.
- EmbeddingGemma-300m — under 200 MB RAM for laptops and phones; 2K context is the limit.
- NVIDIA Nemotron-3-Embed-8B — launched on RTEB rather than MTEB; OpenMDW-1.1 license.
Do you actually need embeddings?
Two 2026 findings should shape the decision:
- Code: Claude Code uses no vector index. Its creator Boris Cherny: “Early versions of Claude Code used RAG + a local vector db, but we found pretty quickly that agentic search generally works better.” Is Grep All You Need? (May 14, 2026) found inline grep beat inline vector search for every harness-and-model pair — and that the harness mattered more than retrieval (the same Opus 4.6 scored 93.1% under one harness and 76.7% under another).
- Documents: a 23,088-query financial-retrieval study found BM25 (nDCG@10 0.515) beat dense retrieval with text-embedding-3-large (0.466); hybrid beat both (0.551); hybrid plus reranking won (0.683). HyDE hurt; Corrective RAG lost to single-shot hybrid.
Where Anthropic, NVIDIA, LlamaIndex and Cursor converged: give the agent a structured-search tool alongside grep and let it choose. Grep wins on small, literal-token corpora; embeddings win where the user says “revenue recognition” and the document says “ASC 606.”
Cost reality check
Ten million tokens a month (~20,000 docs at 512 tokens) costs $0.20 on voyage-4-lite, $1.20 on voyage-4-large, $1.30 on text-embedding-3-large, $2.00 on Gemini Embedding 2. Re-embedding ten million 512-token chunks (5.12B tokens) is $102 on voyage-4-lite, $614 on voyage-4-large, or $412 via Batch. Self-hosting at low volume is not a cost decision — an always-on GPU costs more than the API — but it buys data residency, deprecation immunity and fine-tuning. The real cost of switching is engineering time, index rebuilds and dual-running two indexes through cutover.
How to choose
- Default: Qwen3-Embedding-8B if you can run a GPU; voyage-4-large if you want an API.
- Multimodal corpus: Gemini Embedding 2 (accept the re-embed risk) or Cohere Embed v4 for very long interleaved documents.
- Edge/on-device: EmbeddingGemma or harrier-oss-v1-270m.
- Strict licensing/legal review: IBM Granite (Apache 2.0) or Harrier (MIT).
- Already on OpenAI and retrieval is “fine”: stay, but benchmark hybrid BM25 before spending on a migration.
- Always: evaluate on your own corpus. Leaderboard coverage for the paid flagships is close to zero.