model: jina-embeddings-v5-omni models

[florian-hoenicke]

Hi,

This PR adds the jina-embeddings-v5-omni nano and small base models to MTEB. The text path is parity-verified against the corresponding v5 text models, so the same task routing is used here.

• ModelMeta is filled for both models
• mteb.get_model_meta(...) works for both models
• Results PR: add: jina-embeddings-v5-omni results results#513

Thanks!

[florian-hoenicke]

Follow-up fix pushed in 320263b: JinaV5OmniWrapper now defaults jinaai/jina-embeddings-v5-omni-nano to torch.float32, which restores text-path parity with jinaai/jina-embeddings-v5-text-nano. Verified on A2 with private HF auth via mteb.get_model(...): omni-nano loads fp32 and matched text-nano exactly on retrieval/query and classification/document probes (max_abs_diff=0.0). Small remains unchanged.

[florian-hoenicke]

Thanks, updated the prompt/task dispatch test in c42e7a9 to use MockRetrievalTask().metadata instead of a hand-rolled task metadata object.

[florian-hoenicke]

Follow-up corner-case fix pushed in 9f40546: I found the omni HF remote code was stripping text before tokenization, which made trailing-space inputs differ from the text models. The private HF repos now preserve whitespace and the MTEB metadata pins those fixed revisions. Verified on A2 through mteb.get_model(...): nano 36db6194... and small 8b4f2c44... match their text counterparts on trailing-space probes with max_abs=0.0.

[Samoed]

JinaV5OmniWrapper now defaults jinaai/jina-embeddings-v5-omni-nano to torch.float32, which restores text-path parity

You can just pass this in loader_kwargs without changes in __init__

[florian-hoenicke]

Updated again after the latest private HF repo fix: MTEB now pins nano 6f88a89e... and small 43affca6.... Verified on A2 against the text counterparts on trailing-space probes and all unique STSBenchmark strings: max_abs=0.0, min cosine 1.0 / 0.99999988.

[florian-hoenicke]

Update: the referenced HF repos are now public and ungated:

• jinaai/jina-embeddings-v5-omni-nano: Hub API reports private: false, gated: false, disabled: false
• jinaai/jina-embeddings-v5-omni-small: Hub API reports private: false, gated: false, disabled: false

I also pushed 605cb14 to address the open review feedback: nano fp32 is now passed through loader_kwargs, and the Jina-specific tests were removed. Local checks passed (test_model_meta.py: 1935 passed; ruff on edited files passed).

[Samoed]

I think we would wait until public release

[florian-hoenicke]

The public release is live now: both referenced model repos are publicly accessible and ungated on the Hub.

• https://huggingface.co/jinaai/jina-embeddings-v5-omni-nano
• https://huggingface.co/jinaai/jina-embeddings-v5-omni-small

Hub API currently reports private: false, gated: false, disabled: false for both. Please let me know if you need any other release artifact before merging.