ENH: update models JSON [audio, embedding, llm, rerank]

[XprobeBot]

Types: audio, embedding, llm, rerank
Automated aggregate sync started at 2026-09-07T02:06:15.725Z.

[gemini-code-assist]

Code Review

This pull request updates the model_spec.json file by updating timestamps, modifying the flash-attn dependency version, and removing the model_ability field from several models. The review feedback highlights two main issues: first, the updated flash-attn version contains a local identifier (+cvte1) not available on public PyPI, which will cause installation failures and should be reverted; second, the removal of the model_ability field from jina-embeddings-v5-omni-small and the WeMM-Embedding models (2B, 4B, and 9B) will break multimodal capability discovery and should be restored.

I am having trouble creating individual review comments. Click here to see my feedback.

xinference/model/embedding/model_spec.json (1464)

The package version flash-attn==2.8.3.post1+cvte1 contains a local version identifier (+cvte1). This local version is likely from a custom or internal build on the machine where the automated sync script was run. Since local versions are not available on public PyPI, attempting to install this package will fail for general users with a No matching distribution found error. Please revert this to the standard public version flash-attn==2.8.3.post1.

        "flash-attn==2.8.3.post1"

References

1. When defining Python package dependencies in model specifications, retain the pinned upstream versions from the model's official requirements.txt rather than substituting them with broad minimum ranges, as long as they are valid published releases on PyPI.

xinference/model/embedding/model_spec.json (1692-1696)

The model_ability field for jina-embeddings-v5-omni-small was removed during the automated sync. Since this is an omni model that supports vision, video, and audio modalities, removing this field will prevent the system from correctly reporting and utilizing its multimodal capabilities. Please restore the model_ability field.

xinference/model/embedding/model_spec.json (1917-1920)

The model_ability field for WeMM-Embedding-2B was removed during the automated sync. Since WeMM is a multimodal embedding model supporting vision and video, removing this field will break capability discovery and multimodal input handling. Please restore the model_ability field.

xinference/model/embedding/model_spec.json (1965-1968)

The model_ability field for WeMM-Embedding-4B was removed during the automated sync. Since WeMM is a multimodal embedding model supporting vision and video, removing this field will break capability discovery and multimodal input handling. Please restore the model_ability field.

xinference/model/embedding/model_spec.json (2013-2016)

The model_ability field for WeMM-Embedding-9B was removed during the automated sync. Since WeMM is a multimodal embedding model supporting vision and video, removing this field will break capability discovery and multimodal input handling. Please restore the model_ability field.