jina-embeddings-v5-text-small
Jina AI
Fifth-generation multilingual text embedding model from Jina AI, released 18 February 2026. A 677M-parameter model built on the Qwen3-0.6B-Base backbone, trained in two stages: embedding distillation from Qwen3-Embedding-4B, then four task-specific LoRA adapters (retrieval, text-matching, clustering, classification) fitted on frozen backbone weights. It produces 1024-dimensional embeddings via last-token pooling with Matryoshka truncation down to 32 dimensions, handles sequences up to 32,768 tokens through RoPE with adjusted base frequencies, supports asymmetric retrieval with Query:/Document: prefixes, and is regularised for binary-quantisation robustness. Weights are CC-BY-NC-4.0.
Strengths
- Four task-specific LoRA adapters (retrieval, text-matching, clustering, classification) selectable per request, trained on a frozen backbone
- 32,768-token maximum sequence length with additional long-context training on 1,000-4,096-token documents
- Matryoshka truncation from 1024 down to 32 dimensions, plus binary quantization support aided by a Global Orthogonal Regularizer
- Distilled from the 6x larger Qwen3-Embedding-4B teacher, and beats that teacher on MMTEB pair classification (42.0 vs 26.8) per the vendor
- GGUF quantizations for llama.cpp and MLX builds for Apple Silicon, alongside sentence-transformers and vLLM integrations
- Broad deployment surface: Jina API, Elastic Inference Service, AWS SageMaker, Hugging Face, Docker and air-gapped installs
Caveats
- Weights are CC-BY-NC-4.0 - non-commercial only; commercial use goes through the hosted API or a separate arrangement with Jina AI
- The vendor's own two surfaces disagree on its MMTEB score: the model page says "67.0 average (task-level)" while the Hugging Face card says "67.7 on MMTEB" - neither figure is used in the benchmarks below
- Matryoshka truncation is not free: the vendor states performance "remains strong above 256 dimensions but degrades noticeably below that threshold"
- Retrieval quality depends on using the correct adapter and the asymmetric Query:/Document: prefixes - the model is trained for asymmetric encoding, so misuse costs accuracy
- Language coverage is stated inconsistently: 32 trained / 93 supported languages on the model page versus "119+ languages" on the Hugging Face card
- Vendor-published diagnostics show effective dimensionality of only 70 of 1024 with a 0.300 noise floor and a retrieval recall cliff at 0.566
- Vendor
- Jina AI
- Context Window
- 32,768 tokens
- License
- CC-BY-NC-4.0
- Release Date
- 2026-02-18
- Modalities
- text
| Benchmark | Score | Max |
|---|---|---|
| MTEB English | 71.7 | 100.0 |
| RTEB | 66.8 | 100.0 |
| BEIR | 56.7 | 100.0 |
| LongEmbed | 66.4 | 100.0 |
| MTEB-M retrieval | 64.9 | 100.0 |
Capabilities
Resources
Pricing
Weights are published under CC-BY-NC-4.0 (non-commercial). Hosted access is the Jina Search Foundation API (https://api.jina.ai/v1/embeddings), sold as prepaid token top-ups; Jina's rate-limit table lists the Embedding API as blocked without a key, 100 RPM & 100,000 TPM on a free key, 500 RPM & 2,000,000 TPM on a paid key and 5,000 RPM & 50,000,000 TPM on a premium key. Also offered through Elastic Inference Service, AWS SageMaker, Hugging Face, Docker and air-gapped deployment. Jina publishes no per-token price as page text, so no rate is asserted here.