The generation model gets the demo. The embedding model decides whether RAG finds the right paragraph. explainx.ai's what is an embedding — with examples covers the intuition; the full vector-search guide covers production plumbing. This post is the companion shortlist — same shape as our agent harnesses ranking: top 10 closed-source APIs and top 10 open-source checkpoints.
TL;DR — closed vs open at a glance
| Question | Closed-source API | Open-source / open-weight |
|---|---|---|
| Who runs the model? | The vendor (Voyage/MongoDB, OpenAI, Google, Cohere…) | You — GPU, CPU, edge, or a host you control |
| Can you air-gap it? | No | Yes |
| Typical cost model | Per-token (or batch) metering | Hardware + ops; $0/token at the meter |
| Fine-tune on private pairs? | Rare / vendor-gated | Yes, if the license allows |
| Best for | Fast shipping, managed SLAs, tiny teams | High volume, compliance, customization |
| Start here | Voyage-3-large or OpenAI 3-small | Qwen3-Embedding-0.6B → 4B |
How this list was built
- Quality signal — retrieval-heavy and multilingual MTEB/MMTEB slices as of mid-2026, plus independent RAG bake-offs — not a single STS number.
- Production fit — docs, SDKs, dim truncation (MRL), context length, and pairing with rerankers.
- License / terms clarity — Apache/MIT weighted for open defaults; NC and community licenses called out explicitly.
- Distinctiveness — domain APIs, multimodal, hybrid dense+sparse, edge, MoE — not ten clones.
- Adoption path — what teams can actually wire into LangChain, LlamaIndex, TEI, vLLM, Ollama, or a one-line vendor SDK.
Top 10 closed-source embedding models
1. Voyage AI — voyage-3-large (MongoDB)
The commercial quality leader for pure retrieval. Long context (often cited around 32K), strong domain variants (code, legal, finance), and the model teams reach for when OpenAI embeddings are “fine” but nDCG is still the bottleneck. Voyage is now part of MongoDB’s AI stack, which matters if you’re already on Atlas Vector Search.
Why choose it: maximize retrieval quality from a managed API.
Watch outs: higher $/token than budget options; smaller ecosystem than OpenAI.
2. Google — gemini-embedding-001
Google’s flagship managed text embedding model — frequently near the top of multilingual MTEB tables among APIs, with aggressive pricing and tight fit for Vertex / Gemini stacks.
Why choose it: multilingual quality at API scale inside Google Cloud.
Watch outs: confirm current dims, truncation, and regional availability on the model card before freezing an index.
3. OpenAI — text-embedding-3-large
3072-d default (truncatable via the dimensions parameter), 8K context, deepest third-party integration surface. Still the premium default when your app already speaks OpenAI.
Why choose it: ecosystem + Matryoshka-style dim control without re-embedding.
Watch outs: last major refresh lagged newer Voyage/Gemini/open leaders on public boards — always re-bench on your corpus.
4. Cohere — Embed v4
Enterprise multilingual workhorse with multimodal (text + image) retrieval and a natural pairing with Cohere Rerank in one vendor.
Why choose it: multilingual + multimodal RAG with a single commercial contract.
Watch outs: check input length limits per endpoint; multimodal pricing differs from text-only.
5. OpenAI — text-embedding-3-small
The cheap, reliable “ship today” closed default. Lower dims / cost than 3-large; still good enough for many English FAQ and support corpora.
Why choose it: lowest-friction OpenAI path; great for prototypes and mid-scale RAG.
Watch outs: leave headroom to upgrade (3-large, Voyage, or open Qwen3) once you have an eval set.
6. Voyage — voyage-3 / voyage-code-3 (and lite tiers)
Same family as #1: general voyage-3 for balanced quality/cost, voyage-code-3 when the corpus is repos and diffs, lite tiers when QPS and bill matter more than the last point of nDCG.
Why choose it: domain-tuned commercial embeddings without self-hosting.
Watch outs: don’t mix voyage-code vectors with a general voyage-3 index.
7. Jina Embeddings API
Managed access to Jina’s long-context and multimodal embedders — useful when you want late-chunking / visual-document retrieval without fighting CC-BY-NC weight licenses yourself.
Why choose it: long docs + multimodal on a commercial API.
Watch outs: open weights are often NC; the API is the commercial path.
8. Amazon Titan Text Embeddings V2
The default for teams standardized on AWS Bedrock + OpenSearch / S3 Vectors. Solid English retrieval, IAM-native, and boring in the good way.
Why choose it: AWS-native RAG without a second vendor for embeddings.
Watch outs: less “leaderboard famous” than Voyage/Gemini — validate on your domain.
9. Azure OpenAI embeddings
OpenAI’s embedding models behind Azure’s enterprise boundary — private networking, regional data residency, and procurement that many Fortune 500 stacks already cleared.
Why choose it: OpenAI quality with Azure compliance wrappers.
Watch outs: quota, region, and deployment-name footguns; same model family tradeoffs as OpenAI directly.
10. Mistral Embed (and Snowflake Arctic via Cortex)
Mistral Embed for EU-centric stacks that already use Mistral LLMs. Snowflake arctic-embed via Cortex when the warehouse is the platform and you want embeddings next to the tables.
Why choose it: stay inside an existing Mistral or Snowflake operating model.
Watch outs: narrower ecosystem than OpenAI; treat as platform-default, not universal SOTA.
Closed snapshot
| Rank | Model / API | Best for | Typical dims | Notes |
|---|---|---|---|---|
| 1 | Voyage voyage-3-large | Max retrieval quality | ~1024 (family varies) | MongoDB-aligned |
| 2 | gemini-embedding-001 | Multilingual API | up to ~3072 | Vertex / Gemini |
| 3 | text-embedding-3-large | OpenAI premium | 3072 (truncatable) | Deep integrations |
| 4 | Cohere Embed v4 | Multilingual + multimodal | 256–1536 | Pair with Rerank |
| 5 | text-embedding-3-small | Cheap OpenAI default | 1536 (truncatable) | Prototypes |
| 6 | Voyage-3 / code-3 | Domain / balanced | family-specific | Don’t mix indexes |
| 7 | Jina Embeddings API | Long + multimodal | ~1024 | Commercial vs NC weights |
| 8 | Titan Embeddings V2 | AWS Bedrock | ~1024 | IAM-native |
| 9 | Azure OpenAI embeddings | Enterprise OpenAI | same as OpenAI | Residency / VNet |
| 10 | Mistral Embed / Arctic Cortex | Platform lock-in | vendor-specific | EU / warehouse |
Top 10 open-source embedding models
1. Qwen3-Embedding-8B (Qwen / Alibaba)
The practical open crown. Instruction-aware, 100+ languages, 32K context, Matryoshka dims, Apache 2.0, with matching Qwen3-Reranker models. Widely cited ~70.6 multilingual MTEB-style aggregates.
Why choose it: best default for commercial self-hosted RAG.
Watch outs: 8B + high dims cost GPU and index RAM — truncate via MRL; use the model’s instruction prefixes.
2. KaLM-Embedding-Gemma3-12B (Tencent)
Frequent #1 on MMTEB aggregates (~72.3 in late-2025 snapshots). Ceiling model for bake-offs.
Why choose it: measure the open multilingual upper bound.
Watch outs: ~12B params; Tencent community license — not Apache.
HF: tencent/KaLM-Embedding-Gemma3-12B-2511
3. Qwen3-Embedding-4B
Nearly the 8B story at half the parameters — often the real production deploy target.
Why choose it: best open cost/quality inside the Qwen3 line.
4. llama-embed-nemotron-8b (NVIDIA)
MMTEB neighbor to Qwen3-8B; strong if you already serve on NVIDIA/vLLM.
Watch outs: NVIDIA community / NSCL-style terms — confirm commercial use.
HF: nvidia/llama-embed-nemotron-8b
5. Harrier OSS 0.6B (Microsoft)
0.6B multilingual with MIT — rare license clarity at this size, strong in independent RAG tests.
HF: microsoft/harrier-oss-v1-0.6b
6. Jina Embeddings v5 text-small (and v4 multimodal)
Compact SOTA text (v5) or unified text+image/docs (v4).
Watch outs: many weights are CC-BY-NC-4.0 — use the API (#7 closed) for commercial, or pick Apache/MIT open models instead.
HF: jinaai/jina-embeddings-v5-text-small
7. EmbeddingGemma-300M (Google DeepMind)
Best-in-class under ~500M for multilingual / code slices; MRL 768→128; edge-friendly.
Watch outs: 2048-token context; check Gemma terms.
HF: google/embeddinggemma-300m
8. BGE-M3 (BAAI)
Dense + multi-vector + sparse in one MIT checkpoint; 100+ languages; up to ~8K tokens.
Why choose it: hybrid retrieval without three vendors.
HF: BAAI/bge-m3
9. Qwen3-Embedding-0.6B
Efficient Apache baseline — first index rebuild off MiniLM for most teams.
10. Nomic Embed Text V2 (MoE)
MoE embeddings (~305M active / ~475M total), Apache 2.0, open training artifacts, MRL 768→256.
Watch outs: 512-token context; use search_query: / search_document: prefixes.
HF: nomic-ai/nomic-embed-text-v2-moe
Open snapshot
| Rank | Model | Params | Dim (MRL) | Context | License |
|---|---|---|---|---|---|
| 1 | Qwen3-Embedding-8B | 8B | 4096 (→32+) | 32K | Apache 2.0 |
| 2 | KaLM-Gemma3-12B | ~12B | 3840 (→64) | 32K | Tencent community |
| 3 | Qwen3-Embedding-4B | 4B | 2560 (→32+) | 32K | Apache 2.0 |
| 4 | llama-embed-nemotron-8b | 8B | 4096 | 32K | NVIDIA community |
| 5 | Harrier OSS 0.6B | 0.6B | 1024 | 32K | MIT |
| 6 | Jina v5 text-small | ~0.6B | 1024 (→32) | 32K | CC-BY-NC-4.0 |
| 7 | EmbeddingGemma-300M | 300M | 768 (→128) | 2K | Gemma terms |
| 8 | BGE-M3 | ~568M | 1024 + sparse/multi | 8K | MIT |
| 9 | Qwen3-Embedding-0.6B | 0.6B | 1024 (→32+) | 32K | Apache 2.0 |
| 10 | Nomic Embed Text V2 | MoE ~475M | 768 (→256) | 512 | Apache 2.0 |
Honorable mentions
| Model | Why it still matters |
|---|---|
| all-mpnet-base-v2 / MiniLM | Ubiquitous starters; short context; English prototypes |
| gte-multilingual-base | Solid Alibaba encoder ~305M |
| Snowflake arctic-embed (weights) | Strong English retrieval; also via Cortex (closed #10) |
| Granite Embedding (IBM) | Tiny Apache English embedders |
How to choose without drowning in MTEB
- Closed vs open first — compliance and ops, not leaderboard ego.
- Build a 100–200 query gold set from your docs (embeddings complete guide).
- A/B two closed + one open on Recall@5 / nDCG@10 before freezing dims.
- Match query/document prompts to the model card.
- Add hybrid BM25 before blaming the embedder for SKU misses (hybrid search guide).
- Rerank top-k when precision pays (Cohere Rerank, Qwen3-Reranker, or a cross-encoder).
# Practical open starting point
huggingface-cli download Qwen/Qwen3-Embedding-0.6B
Same pattern as open vs closed agent harnesses and open-weight vs closed models: prototype where shipping is cheap, standardize where control and unit economics matter.
Related on explainx.ai
- What is an embedding? Examples + interactive demo
- What are embeddings? Complete vector-search guide
- Semantic vs vector vs hybrid search
- Prompt engineering vs fine-tuning vs RAG
- Grounding: RAG vs fine-tuning decision guide
- Top 10 open and closed agent harnesses
- Ternlight browser WASM embeddings
Official starting points
Ranks, pricing, dims, and licenses drift — re-check vendor docs and model cards before you rebuild an index. Snapshot as of July 28, 2026.