Skip to main content

Log in to comment and vote

Comments2

  • Rose Astrometry

    •

    Mar 9

    I would like to specifically request that the vision-capable embedding model “Qwen3-VL-Embedding” be added to the Venice API. Please do NOT merely add the non-vision variant because the VL variant can do so much more! Three critical features make this embedding model stand out from the crowd:

    1) The model maps text and images (and video) embeddings into the same embedding space. This feature enables text-to-image search.

    2) The model accepts multiple images as inputs. That means the model can create embeddings that represent more than can be shown in a single image. For example, embedding several photos of an object shown from multiple angles results in much more rich embeddings than those generated from a single photo. 

    3) The model follows instructions! Unlike most other embedding models, the embeddings generated by this model can be instructed to represent specific features contained in the input. For example, instructing the model to “Represent the colors in this image” will result in different embeddings than instructing the model to “Represent the objects in this image”. The “color” embeddings enable search or clustering images by the colors shown in the images. The “objects” embeddings enable search or clustering images by objects shown in the images.

    We have built multiple apps that utilize Qwen3-VL-Embedding in production, but we currently source our embeddings elsewhere.

    Please do add the VL variant to the Venice API asap :)

  • An Anonymous User

    •

    Dec 29, 2025

    •

    Merged request

    •

    1 vote

    Add Qwen 3 embedding (0.6B, 4B, 8B) models