AI Potluck
Back to Gap Map Model components / Embeddings & retrieval models

Jina Reranker

Jina AI
restricted / Overall score: 4.0(strong)

Jina AI's reranker line, from the v1 turbo and tiny English cross-encoders through v2-base-multilingual to the v3 and v3.5 listwise rerankers. v3 is a 0.6B Qwen3-derived model with a "last but not late interaction" architecture that attends across query and up to 64 documents inside one 131K-token window, reporting 61.94 nDCG@10 on BEIR. The weights are published on the Hub, but under a non-commercial licence.

Added 2026-09-11 for the embeddings_retrieval promotion. Scope is the TEXT line: the m0 multimodal reranker is excluded on every axis under the category's rule that a vision-language line is a separate product from the text line, matching how jina-embeddings treats v4 and v5-omni. The licence is the thing to read carefully - the v2, v3 and v3.5 cards all read cc-by-nc-4.0 and v3's own licence section says commercial use outside AWS and Azure needs a separate conversation, while the older v1-turbo-en and v1-tiny-en are Apache-2.0. The most restrictive distributed SKU governs. Verified 2026-09-11 via the jina-reranker-v3 model card.

Openness

2 high confidence
2.0
weights
open(ungated safetensors on the Hub for v1 through v3.5)
data
closed(no training data or construction scripts released)
code
partial(remote inference code shipped with the card
license
Apache-2.0(OSI, and only the superseded v1-turbo-en and v1-tiny-en)+CC-BY-NC-4.0(non-commercial

The current SKUs are licensed CC-BY-NC-4.0 and the superseded v1-turbo-en and v1-tiny-en are Apache-2.0, so the compound resolves on the non-commercial half. commercial_forbidden asks one question - does the licence permit commercial use at all - and the card answers no, saying commercial use beyond AWS and Azure requires contacting sales. The m0 multimodal reranker is excluded from this record on every axis under the category's vision-language rule; it reads cc-by-nc-4.0 as well, so its exclusion moves nothing here.

Adoption

4 high confidence
4.0

About 2.0M downloads in the trailing 30 days across the text line - v2-base-multilingual 1.16M, v3 780K, v3.5 32K, plus the two v1 checkpoints. The v2 and v3 figures alone clear the 1M band. The m0 multimodal reranker is excluded under the category's rule that a vision-language line is a separate product from the text line; counting its 379K gives about 2.4M and the same band.

Capability

4 high confidence
4.0

Rung 4, competitive frontier. Level with qwen3-reranker: 61.94 BEIR against Qwen3-Reranker-4B at 61.16 on the same table, at a sixth of the parameters, and multilingual with a long context in both cases. Note the numbers come from the publisher own card, which is the instrument available for this product and is why the comparison is drawn against a peer scored from the same table.

  • https://huggingface.co/jinaai/jina-reranker-v3/raw/main/README.md recorded 2026-09-11

    Card reports jina-reranker-v3 at 61.94 BEIR, 66.83 MIRACL, 67.92 MKQA and 70.64 CoIR at 0.6B parameters, in a table that also lists Qwen3-Reranker-4B at 61.16 BEIR and bge-reranker-v2-m3 at 56.51; describes 64 documents within a 131K context window.

Verified 2026-09-11