Jina Reranker v3.5
Paid production API
Type / status: rerank / documented_api
Use cases: RAG, Document reranking
Supported languages: multilingual
Deployment options: elastic, api, huggingface, aws, azure, gcp, airgapped
Documented datacenters: US
Model release date: 2026-07-27
API endpoint: https://api.jina.ai/v1/rerank
Open-weights license (hosted API terms are separate): CC-BY-NC-4.0
Official catalog input-token price. Embeddings produce vectors and rerankers produce rankings, not billable generated output tokens. Zero catalog completion/image/request values do not make the service free. Reader API billing is separate from the model-catalog price.
The Reader page advertises 10 million free tokens for a new API key. Actual allocation and expiry are subject to the official account; this is a trial, not permanent free access.
Subject to hosted service terms; buying API tokens does not license model weights. Each model has its own weight license. CC-BY-NC-4.0 and Qwen Research License must not be described as unrestricted commercial self-hosting. Official FAQ states API requests, inputs and outputs are not used to train models. Jina legal page directs current data processing to Elastic's Data Processing Addendum. Model catalog lists US datacenters; Reader/Search also document EU endpoints. Do not infer identical regions across products.
Official model ID: jina-ai/jina-reranker-v3.5
Input / output: text → ranking
Model developer: Jina AI · Service provider: Jina AI
Model family: — · Currency: USD
| Price dimension | Current rate |
|---|---|
| Input | USD 0.05 |
| Cached read | Unverified |
jina-reranker-v3.5 is a 0.6B parameter multilingual listwise document reranker and a drop-in upgrade to jina-reranker-v3. It keeps the same last-but-not-late (LBNL) listwise interface — ranking a query against many candidates in a single forward pass — while improving domain robustness, structured-data ranking, multilingual retrieval, and inference efficiency. Built on Qwen3-0.6B with 28 layers using hybrid 3L2G attention (sliding-window layers, window 1024, interleaved with global-attention layers and a pinned terminal global layer required by LBNL) and a lightweight MLP projector (1024→512→512), it supports up to 131K tokens of context and ranks documents jointly via causal self-attention and cosine scoring. It reaches 63.20 nDCG@10 on BEIR — above the 4B Qwen3-Reranker at roughly 7× fewer parameters — while running 1.22×–1.56× faster than v3 across short-to-long contexts. Endpoint and model identifier are declared in the current OpenAPI schema; authenticated inference was not executed. Endpoint support is documented; authenticated inference was not tested. The output is ranking scores, not generated text tokens.
Documented API context: 131,072 tokens
Official source · Last verified: 2026-10-15T09:33:13.684489+00:00