Embedding and reranking models (2026-09-30)
Overview of current embedding models from vendor documentation. Scores are omitted unless vendor-cited. Embedding spaces are not interchangeable: changing model means re-embedding your corpus.
| Model | Type | Key specs | Licence / access |
|---|---|---|---|
OpenAI text-embedding-3-small / -large | Text | 1,536 / 3,072 dims, 8,192 tokens; dimensions parameter for truncation | Hosted API. No newer OpenAI embedding model appears in the docs fetched |
Google gemini-embedding-2 | Multimodal (text, image, video, audio, PDF) | 8,192 tokens; 128-3,072 dims (768/1536/3072 recommended); listed as stable, updated April 2026 (as of 2026-09-30 docs) | Gemini API |
Google gemini-embedding-001 | Text | 2,048 tokens; 128-3,072 dims; incompatible space with v2 | Gemini API |
| embeddinggemma | Text, on-device | 300M params, 768 dims (MRL to 128), 2,048 tokens, 100+ languages (2025-09-24) | Open weights, Gemma terms |
Cohere embed-v5.0-pro / -fast | Multimodal (text, images) | 128k-token context; dims 256-2048 | Hosted API; v4.0 and v3.0 remain |
Voyage voyage-4-large / voyage-4 / voyage-4-lite, voyage-code-4 | Text, code | 32,000 tokens; dims 256/512/1024/2048; int8/binary quantisation | Hosted; voyage-4-nano open weights on Hugging Face |
Jina jina-embeddings-v5-text (small 677M, nano 239M); v5-omni multimodal | Text / multimodal | 32K / 8K context | CC BY-NC 4.0; per jina.ai (2026-09-30) commercial licence via Elastic (SKU from 2026-08-10) and Jina is part of Elastic (not checked against an Elastic primary source) |
| Qwen3-Embedding (0.6B / 4B / 8B) | Text, open | 8B: 32k context, up to 4096 dims, 100+ languages, instruction-aware; MTEB multilingual 70.58 (June 2025) | Apache 2.0 |
| BAAI BGE-M3 | Text, open | 1024 dims, 8,192 tokens, dense + sparse + multi-vector | MIT |
Rerankers
- Cohere Rerank (current version not checked).
- Jina
jina-reranker-v3and v3.5 (names not verified). - Qwen3-Reranker 0.6B / 4B / 8B (licence not checked).
- Voyage rerankers exist; versions unverified.
Notes
- Moved from DRAFT to OK on 2026-10-05: the table was read from vendor docs on 2026-09-30 and not re-fetched since; it is a snapshot, so re-check model names before relying on it.
- Several hosted models list multimodal inputs per their docs as of 2026-09-30 (Gemini Embedding 2, Cohere v5, Jina omni).
- Check MTEB/MMTEB leaderboards for up-to-date rankings; none were verified here.
Related
agent-memory-systems, graphrag, whisper-and-asr-models
Open items
- Not covered or not verified: Nvidia, Mistral, IBM Granite and Snowflake embedding models; current Cohere Rerank, Voyage reranker and Qwen3-Reranker versions and licences; MTEB/MMTEB rankings. Prices deliberately omitted.
Sources
- https://ai.google.dev/gemini-api/docs/embeddings (accessed 2026-09-30)
- https://developers.openai.com/api/docs/guides/embeddings (accessed 2026-09-30)
- https://docs.cohere.com/docs/cohere-embed (accessed 2026-09-30)
- https://docs.voyageai.com/docs/embeddings (accessed 2026-09-30)
- https://jina.ai/embeddings/ (accessed 2026-09-30)
- https://huggingface.co/Qwen/Qwen3-Embedding-8B (accessed 2026-09-30)
- https://huggingface.co/BAAI/bge-m3 (accessed 2026-09-30)
- https://huggingface.co/google/embeddinggemma-300m (accessed 2026-09-30)